doi
stringlengths 9
10
| chunk-id
stringlengths 1
4
| chunk
stringlengths 1
1.57k
| id
stringlengths 9
10
| title
stringlengths 10
127
| summary
stringlengths 581
1.92k
| source
stringlengths 30
31
| authors
sequence | categories
sequence | comment
stringclasses 28
values | journal_ref
stringclasses 1
value | primary_category
stringclasses 6
values | published
stringlengths 8
8
| updated
stringlengths 8
8
| references
list |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
1906.04016 | 0 | Learning Temporal Pose Estimation
from Sparsely-Labeled Videos
Gedas Bertasius1;2, Christoph Feichtenhofer1, Du Tran1, Jianbo Shi2, Lorenzo Torresani1
1Facebook AI,2University of Pennsylvania
Abstract
Modern approaches for multi-person pose estimation in video require large amounts
of dense annotations. However, labeling every frame in a video is costly and labor
intensive. To reduce the need for dense annotations, we propose a PoseWarper
network that leverages training videos with sparse annotations (every kframes) to
learn to perform dense temporal pose propagation and estimation. Given a pair
of video frames—a labeled Frame A and an unlabeled Frame B—we train our
model to predict human pose in Frame A using the features from Frame B by
means of deformable convolutions to implicitly learn the pose warping between A
and B. We demonstrate that we can leverage our trained PoseWarper for several
applications. First, at inference time we can reverse the application direction
of our network in order to propagate pose information from manually annotated
frames to unlabeled frames. This makes it possible to generate pose annotations
for the entire video given only a few manually-labeled frames. Compared to
modern label propagation methods based on optical flow, our warping mechanism | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 1 | for the entire video given only a few manually-labeled frames. Compared to
modern label propagation methods based on optical flow, our warping mechanism
is much more compact ( 6M vs 39M parameters), and also more accurate ( 88:7%
mAP vs 83:8%mAP). We also show that we can improve the accuracy of a
pose estimator by training it on an augmented dataset obtained by adding our
propagated poses to the original manual labels. Lastly, we can use our PoseWarper
to aggregate temporal pose information from neighboring frames during inference.
This allows us to obtain state-of-the-art pose detection results on PoseTrack2017
and PoseTrack2018 datasets. Code has been made available at: https://github.
com/facebookresearch/PoseWarper .
1 Introduction
In recent years, visual understanding methods [ 1–15] have made tremendous progress, partly because
of advances in deep learning [ 16–19], and partly due to the introduction of large-scale annotated
datasets [ 20,21]. In this paper we consider the problem of pose estimation, which has greatly
benefitted from the recent creation of large-scale datasets [ 22,23]. Most of the recent advances
in this area, though, have concentrated on the task of pose estimation in still-images [ 3,23–27]. | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 2 | in this area, though, have concentrated on the task of pose estimation in still-images [ 3,23–27].
However, directly applying these image-level models to video is challenging due to nuisance factors
such as motion blur, video defocus, and frequent pose occlusions. Additionally, the process of
collecting annotated pose data in multi-person videos is costly and time consuming. A video typically
contains hundreds of frames that need to be densely-labeled by human annotators. As a result,
datasets for video pose estimation [ 22] are typically smaller and less diverse compared to their image
counterparts [ 21]. This is problematic because modern deep models require large amounts of labeled
data to achieve good performance. At the same time, videos have high informational redundancy as
the content changes little from frame to frame. This raises the question of whether every single frame
in a training video needs to be labeled in order to achieve good pose estimation accuracy.
33rd Conference on Neural Information Processing Systems (NeurIPS 2019), Vancouver, Canada.arXiv:1906.04016v3 [cs.CV] 11 Dec 2019
To reduce the reliance on densely annotated video pose data, in this work, we introduce the PoseWarper network, which operates on sparsely annotated videos, i.e., videos where pose annotations | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 3 | are given only every kframes. Given a pair of frames from the same video—a labeled Frame A
and an unlabeled Frame B—we train our model to detect pose in Frame A, using the features from
Frame B. To achieve this goal, our model leverages deformable convolutions [ 28] across space and
time. Through this mechanism, our model learns to sample features from an unlabeled Frame B to
maximize pose detection accuracy in a labeled Frame A.
Our trained PoseWarper can then be used for several applications. First, we can leverage PoseWarper
to propagate pose information from a few manually-labeled frames across the entire video. Compared
to modern optical flow propagation methods such as FlowNet2 [ 29], our PoseWarper produces more
accurate pose annotations ( 88:7%mAP vs 83:8%mAP), while also employing a much more compact
warping mechanism ( 6M vs 39M parameters). Furthermore, we show that our propagated poses can
serve as effective pseudo labels for training a more accurate pose detector. Finally, our PoseWarper
can be used to aggregate temporal pose information from neighboring frames during inference. This
naturally renders the approach more robust to occlusion or motion blur in individual frames, and leads
to state-of-the-art pose detection results on the PoseTrack2017 and PoseTrack2018 datasets [22]. | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 4 | to state-of-the-art pose detection results on the PoseTrack2017 and PoseTrack2018 datasets [22].
2 Related Work
Multi-Person Pose Detection in Images. The traditional approaches for pose estimation leverage
pictorial structures model [ 30–34], which represents human body as a tree-structured graph with
pairwise potentials between the connected body parts. These approaches have been highly successful
in the past, but they tend to fail if some of body parts are occluded. These issues have been partially
addressed by the models that assume a non-tree graph structure [ 35–38]. However, most modern
approaches for single image pose estimation are based on convolutional neural networks [ 3,6,23–
27,39–45]. The method in [ 3] regresses (x;y)joint coordinates directly from the images. More
recent work [ 25] instead predicts pose heatmaps, which leads to an easier optimization problem.
Several approaches [ 24,26,39,46] propose an iterative pose estimation pipeline where the predictions
are refined at different stages inside a CNN or via a recurrent network. The methods in [ 6,23,45]
tackle pose estimation problem in a top-down fashion, first detecting bounding boxes of people, and | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 5 | tackle pose estimation problem in a top-down fashion, first detecting bounding boxes of people, and
then predicting the pose heatmaps from the cropped images. The work in [ 24] proposes part affinity
fields module that captures pairwise relationships between different body parts. The approaches
in [42,43] leverage a bottom-up pipeline first predicting the keypoints, and then assembling them
into instances. Lastly, a recent work in [ 27], proposes an architecture that preserves high resolution
feature maps, which is shown to be highly beneficial for the multi-person pose estimation task.
Multi-Person Pose Detection in Video. Due to a limited number of large scale benchmarks for
video pose detection, there has been significantly fewer methods in the video domain. Several prior
methods [ 22,47,48] tackle a video pose estimation task as a two-stage problem, first detecting the
keypoints in individual frames, and then applying temporal smoothing techniques. The method in [ 49]
proposes a spatiotemporal CRF, which is jointly optimized for the pose prediction in video. The
work in [ 50] proposes a personalized video pose estimation framework, which is accomplished by | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 6 | proposes a spatiotemporal CRF, which is jointly optimized for the pose prediction in video. The
work in [ 50] proposes a personalized video pose estimation framework, which is accomplished by
finetuning the model on a few frames with high confidence keypoints in each video. The approaches
in [51,52] leverage flow based representations for aligning features temporally across multiple
frames, and then using such aligned features for pose detection in individual frames.
In contrast to these prior methods, our primary objective is to learn an effective video pose detector
from sparsely labeled videos. Our approach has similarities to the methods in [ 51,52], which use
flow representations for feature alignment. However, unlike our model, the methods in [ 51,52] do
not optimize their flow representations end-to-end with respect to the pose detection task. As we will
show in our experiments, this is important for strong performance.
3 The PoseWarper Network
Overview. Our goal is to design a model that learns to detect pose from sparsely labeled videos.
Specifically, we assume that pose annotations in training videos are available every kframes. Inspired
by a recent self-supervised approach for learning facial attribute embeddings [ 53], we formulate the | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 7 | by a recent self-supervised approach for learning facial attribute embeddings [ 53], we formulate the
following task. Given two video frames—a labeled Frame A and an unlabeled Frame B—our model
2
1. Training (Unlabeled Labeled)
2. Video Pose Propagation (Labeled Unlabeled) | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 8 | Unlabeled Frame B (time t+ ) <latexit | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 9 | sha1_base64="AU/vq+DAgMm1x9RhUc20aA7BJwY=">AAAB7XicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cK9gPaUDabTbt2kw27E6GE/gcvHhTx6v/x5r9x2+agrQ8GHu/NMDMvSKUw6LrfTmltfWNzq7xd2dnd2z+oHh61jco04y2mpNLdgBouRcJbKFDybqo5jQPJO8H4duZ3nrg2QiUPOEm5H9NhIiLBKFqp3Q+5RDqo1ty6OwdZJV5BalCgOah+9UPFspgnyCQ1pue5Kfo51SiY5NNKPzM8pWxMh | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 10 | ah+9UPFspgnyCQ1pue5Kfo51SiY5NNKPzM8pWxMh7xnaUJjbvx8fu2UnFklJJHSthIkc/X3RE5jYyZxYDtjiiOz7M3E/7xehtG1n4skzZAnbLEoyiRBRWavk1BozlBOLKFMC3srYSOqKUMbUMWG4C2/vEraF3XPrXv3l7XGTRFHGU7gFM7BgytowB00oQUMHuEZXuHNUc6L8+58LFpLTjFzDH/gfP4AkYyPHA==</latexit><latexit | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 11 | sha1_base64="AU/vq+DAgMm1x9RhUc20aA7BJwY=">AAAB7XicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cK9gPaUDabTbt2kw27E6GE/gcvHhTx6v/x5r9x2+agrQ8GHu/NMDMvSKUw6LrfTmltfWNzq7xd2dnd2z+oHh61jco04y2mpNLdgBouRcJbKFDybqo5jQPJO8H4duZ3nrg2QiUPOEm5H9NhIiLBKFqp3Q+5RDqo1ty6OwdZJV5BalCgOah+9UPFspgnyCQ1pue5Kfo51SiY5NNKPzM8pWxMh | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 12 | ah+9UPFspgnyCQ1pue5Kfo51SiY5NNKPzM8pWxMh7xnaUJjbvx8fu2UnFklJJHSthIkc/X3RE5jYyZxYDtjiiOz7M3E/7xehtG1n4skzZAnbLEoyiRBRWavk1BozlBOLKFMC3srYSOqKUMbUMWG4C2/vEraF3XPrXv3l7XGTRFHGU7gFM7BgytowB00oQUMHuEZXuHNUc6L8+58LFpLTjFzDH/gfP4AkYyPHA==</latexit><latexit | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 13 | sha1_base64="AU/vq+DAgMm1x9RhUc20aA7BJwY=">AAAB7XicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cK9gPaUDabTbt2kw27E6GE/gcvHhTx6v/x5r9x2+agrQ8GHu/NMDMvSKUw6LrfTmltfWNzq7xd2dnd2z+oHh61jco04y2mpNLdgBouRcJbKFDybqo5jQPJO8H4duZ3nrg2QiUPOEm5H9NhIiLBKFqp3Q+5RDqo1ty6OwdZJV5BalCgOah+9UPFspgnyCQ1pue5Kfo51SiY5NNKPzM8pWxMh | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 14 | ah+9UPFspgnyCQ1pue5Kfo51SiY5NNKPzM8pWxMh7xnaUJjbvx8fu2UnFklJJHSthIkc/X3RE5jYyZxYDtjiiOz7M3E/7xehtG1n4skzZAnbLEoyiRBRWavk1BozlBOLKFMC3srYSOqKUMbUMWG4C2/vEraF3XPrXv3l7XGTRFHGU7gFM7BgytowB00oQUMHuEZXuHNUc6L8+58LFpLTjFzDH/gfP4AkYyPHA==</latexit><latexit | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 15 | sha1_base64="AU/vq+DAgMm1x9RhUc20aA7BJwY=">AAAB7XicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cK9gPaUDabTbt2kw27E6GE/gcvHhTx6v/x5r9x2+agrQ8GHu/NMDMvSKUw6LrfTmltfWNzq7xd2dnd2z+oHh61jco04y2mpNLdgBouRcJbKFDybqo5jQPJO8H4duZ3nrg2QiUPOEm5H9NhIiLBKFqp3Q+5RDqo1ty6OwdZJV5BalCgOah+9UPFspgnyCQ1pue5Kfo51SiY5NNKPzM8pWxMh | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 16 | ah+9UPFspgnyCQ1pue5Kfo51SiY5NNKPzM8pWxMh7xnaUJjbvx8fu2UnFklJJHSthIkc/X3RE5jYyZxYDtjiiOz7M3E/7xehtG1n4skzZAnbLEoyiRBRWavk1BozlBOLKFMC3srYSOqKUMbUMWG4C2/vEraF3XPrXv3l7XGTRFHGU7gFM7BgytowB00oQUMHuEZXuHNUc6L8+58LFpLTjFzDH/gfP4AkYyPHA==</latexit>Labeled | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 17 | Frame A (time t) | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 18 | Features Relating the Two Frames
Predicted Pose in Frame A
Forward Pass
Backward Pass | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 19 | Labeled Frame A (time t+ ) <latexit | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 20 | sha1_base64="AU/vq+DAgMm1x9RhUc20aA7BJwY=">AAAB7XicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cK9gPaUDabTbt2kw27E6GE/gcvHhTx6v/x5r9x2+agrQ8GHu/NMDMvSKUw6LrfTmltfWNzq7xd2dnd2z+oHh61jco04y2mpNLdgBouRcJbKFDybqo5jQPJO8H4duZ3nrg2QiUPOEm5H9NhIiLBKFqp3Q+5RDqo1ty6OwdZJV5BalCgOah+9UPFspgnyCQ1pue5Kfo51SiY5NNKPzM8pWxMh | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 21 | ah+9UPFspgnyCQ1pue5Kfo51SiY5NNKPzM8pWxMh7xnaUJjbvx8fu2UnFklJJHSthIkc/X3RE5jYyZxYDtjiiOz7M3E/7xehtG1n4skzZAnbLEoyiRBRWavk1BozlBOLKFMC3srYSOqKUMbUMWG4C2/vEraF3XPrXv3l7XGTRFHGU7gFM7BgytowB00oQUMHuEZXuHNUc6L8+58LFpLTjFzDH/gfP4AkYyPHA==</latexit><latexit | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 22 | sha1_base64="AU/vq+DAgMm1x9RhUc20aA7BJwY=">AAAB7XicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cK9gPaUDabTbt2kw27E6GE/gcvHhTx6v/x5r9x2+agrQ8GHu/NMDMvSKUw6LrfTmltfWNzq7xd2dnd2z+oHh61jco04y2mpNLdgBouRcJbKFDybqo5jQPJO8H4duZ3nrg2QiUPOEm5H9NhIiLBKFqp3Q+5RDqo1ty6OwdZJV5BalCgOah+9UPFspgnyCQ1pue5Kfo51SiY5NNKPzM8pWxMh | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 23 | ah+9UPFspgnyCQ1pue5Kfo51SiY5NNKPzM8pWxMh7xnaUJjbvx8fu2UnFklJJHSthIkc/X3RE5jYyZxYDtjiiOz7M3E/7xehtG1n4skzZAnbLEoyiRBRWavk1BozlBOLKFMC3srYSOqKUMbUMWG4C2/vEraF3XPrXv3l7XGTRFHGU7gFM7BgytowB00oQUMHuEZXuHNUc6L8+58LFpLTjFzDH/gfP4AkYyPHA==</latexit><latexit | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 24 | sha1_base64="AU/vq+DAgMm1x9RhUc20aA7BJwY=">AAAB7XicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cK9gPaUDabTbt2kw27E6GE/gcvHhTx6v/x5r9x2+agrQ8GHu/NMDMvSKUw6LrfTmltfWNzq7xd2dnd2z+oHh61jco04y2mpNLdgBouRcJbKFDybqo5jQPJO8H4duZ3nrg2QiUPOEm5H9NhIiLBKFqp3Q+5RDqo1ty6OwdZJV5BalCgOah+9UPFspgnyCQ1pue5Kfo51SiY5NNKPzM8pWxMh | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 25 | ah+9UPFspgnyCQ1pue5Kfo51SiY5NNKPzM8pWxMh7xnaUJjbvx8fu2UnFklJJHSthIkc/X3RE5jYyZxYDtjiiOz7M3E/7xehtG1n4skzZAnbLEoyiRBRWavk1BozlBOLKFMC3srYSOqKUMbUMWG4C2/vEraF3XPrXv3l7XGTRFHGU7gFM7BgytowB00oQUMHuEZXuHNUc6L8+58LFpLTjFzDH/gfP4AkYyPHA==</latexit><latexit | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 26 | sha1_base64="AU/vq+DAgMm1x9RhUc20aA7BJwY=">AAAB7XicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cK9gPaUDabTbt2kw27E6GE/gcvHhTx6v/x5r9x2+agrQ8GHu/NMDMvSKUw6LrfTmltfWNzq7xd2dnd2z+oHh61jco04y2mpNLdgBouRcJbKFDybqo5jQPJO8H4duZ3nrg2QiUPOEm5H9NhIiLBKFqp3Q+5RDqo1ty6OwdZJV5BalCgOah+9UPFspgnyCQ1pue5Kfo51SiY5NNKPzM8pWxMh | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 27 | ah+9UPFspgnyCQ1pue5Kfo51SiY5NNKPzM8pWxMh7xnaUJjbvx8fu2UnFklJJHSthIkc/X3RE5jYyZxYDtjiiOz7M3E/7xehtG1n4skzZAnbLEoyiRBRWavk1BozlBOLKFMC3srYSOqKUMbUMWG4C2/vEraF3XPrXv3l7XGTRFHGU7gFM7BgytowB00oQUMHuEZXuHNUc6L8+58LFpLTjFzDH/gfP4AkYyPHA==</latexit>Unlabeled | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 28 | Frame B (time t) | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 29 | Ground Truth Pose Propagated to Frame BFeature Warping
Ground Truth Pose in Frame AFeatures Relating the Two FramesPose Detection
Feature WarpingFigure 1: A high level overview of our approach for using sparsely labeled videos for pose detection.
Faces in the figure are artificially masked for privacy reasons. In each training video, pose annotations
are available only every kframes. During training, our system considers a pair of frames–a labeled
Frame A, and an unlabeled Frame B, and aims to detect pose in Frame A, using the features from
Frame B. Our training procedure is designed to achieve two goals: 1) our model must be able to
extract motion offsets relating these two frames. 2) Using these motion offsets our model must
then be able to rewarp the detected pose heatmap extracted from an unlabeled Frame B in order to
optimize the accuracy of pose detection in a labeled Frame A. After training, we can apply our model
in reverse order to propagate pose information across the entire video from ground truth poses given
for only a few frames.
is allowed to compare Frame A to Frame B but it must predict Pose A (i.e., the pose in Frame A)
using the features from Frame B, as illustrated in Figure 1 (top). | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 30 | using the features from Frame B, as illustrated in Figure 1 (top).
At first glance, this task may look overly challenging: how can we predict Pose A by merely using
features from Frame B? However, suppose that we had body joint correspondences between Frame A
and Frame B. In such a scenario, this task would become trivial, as we would simply need to spatially
“warp” the feature maps computed from frame B according to the set of correspondences relating
frame B to frame A. Based on this intuition, we design a learning scheme that achieves two goals: 1)
By comparing Frame A and Frame B, our model must be able to extract motion offsets relating these
two frames. 2) Using these motion offsets our model must be able to rewarp the pose extracted from
an unlabeled Frame B in order to optimize pose detection accuracy in a labeled Frame A.
To achieve these goals, we first feed both frames through a backbone CNN that predicts pose heatmaps
for each of the frames. Then, the resulting heatmaps from both frames are used to determine which
points from Frame B should be sampled for detection in Frame A. Finally, the resampled pose
heatmap from Frame B is used to maximize accuracy of Pose A. | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 31 | points from Frame B should be sampled for detection in Frame A. Finally, the resampled pose
heatmap from Frame B is used to maximize accuracy of Pose A.
Backbone Network. Due to its high efficiency and accuracy, we use the state-of-the-art High
Resolution Network (HRNet-W48) [ 27] as our backbone CNN. However, we note that our system can
easily integrate other architectures as well. Thus, we envision that future improvements in still-image
pose estimation will further improve the effectiveness of our approach.
Deformable Warping. Initially, we feed Frame A and Frame B through our backbone CNN, which
outputs pose heatmaps fAandfB. Then, we compute the difference A;B=fA fB. The resulting
feature tensor A;Bis provided as input to a stack of 33simple residual blocks (as in standard
ResNet-18 or ResNet-34 models), which output a feature tensor A;B. The feature tensor A;Bis
then fed into five 33convolutional layers each using a different dilation rate d2f3;6;12;18;24g
3
Labeled Frame A (time t)
Offsets | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 32 | Warped Feature MapDifference <latexit | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 33 | sha1_base64="xbJVLRy5zV8jtLsjAis5DXwMT4U=">AAAB7nicbVDJSgNBEK1xjXGLevTSGARPYUYEPQa9eIxgFkiG0NPpSZr0MvQihCEf4cWDIl79Hm/+jZ1kDpr4oODxXhVV9ZKMM2PD8DtYW9/Y3Nou7ZR39/YPDitHxy2jnCa0SRRXupNgQzmTtGmZ5bSTaYpFwmk7Gd/N/PYT1YYp+WgnGY0FHkqWMoKtl9o9JZh0pl+phrVwDrRKooJUoUCjX/nqDRRxgkpLODamG4WZjXOsLSOcTss9Z2iGyRgPa | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 34 | /nqDRRxgkpLODamG4WZjXOsLSOcTss9Z2iGyRgPaddTiQU1cT4/d4rOvTJAqdK+pEVz9fdEjoUxE5H4ToHtyCx7M/E/r+tsehPnTGbOUkkWi1LHkVVo9jsaME2J5RNPMNHM34rICGtMrE+o7EOIll9eJa3LWhTWooerav22iKMEp3AGFxDBNdThHhrQBAJjeIZXeAuy4CV4Dz4WrWtBMXMCfxB8/gCVCY+3</latexit><latexit | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 35 | sha1_base64="xbJVLRy5zV8jtLsjAis5DXwMT4U=">AAAB7nicbVDJSgNBEK1xjXGLevTSGARPYUYEPQa9eIxgFkiG0NPpSZr0MvQihCEf4cWDIl79Hm/+jZ1kDpr4oODxXhVV9ZKMM2PD8DtYW9/Y3Nou7ZR39/YPDitHxy2jnCa0SRRXupNgQzmTtGmZ5bSTaYpFwmk7Gd/N/PYT1YYp+WgnGY0FHkqWMoKtl9o9JZh0pl+phrVwDrRKooJUoUCjX/nqDRRxgkpLODamG4WZjXOsLSOcTss9Z2iGyRgPa | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 36 | /nqDRRxgkpLODamG4WZjXOsLSOcTss9Z2iGyRgPaddTiQU1cT4/d4rOvTJAqdK+pEVz9fdEjoUxE5H4ToHtyCx7M/E/r+tsehPnTGbOUkkWi1LHkVVo9jsaME2J5RNPMNHM34rICGtMrE+o7EOIll9eJa3LWhTWooerav22iKMEp3AGFxDBNdThHhrQBAJjeIZXeAuy4CV4Dz4WrWtBMXMCfxB8/gCVCY+3</latexit><latexit | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 37 | sha1_base64="xbJVLRy5zV8jtLsjAis5DXwMT4U=">AAAB7nicbVDJSgNBEK1xjXGLevTSGARPYUYEPQa9eIxgFkiG0NPpSZr0MvQihCEf4cWDIl79Hm/+jZ1kDpr4oODxXhVV9ZKMM2PD8DtYW9/Y3Nou7ZR39/YPDitHxy2jnCa0SRRXupNgQzmTtGmZ5bSTaYpFwmk7Gd/N/PYT1YYp+WgnGY0FHkqWMoKtl9o9JZh0pl+phrVwDrRKooJUoUCjX/nqDRRxgkpLODamG4WZjXOsLSOcTss9Z2iGyRgPa | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 38 | /nqDRRxgkpLODamG4WZjXOsLSOcTss9Z2iGyRgPaddTiQU1cT4/d4rOvTJAqdK+pEVz9fdEjoUxE5H4ToHtyCx7M/E/r+tsehPnTGbOUkkWi1LHkVVo9jsaME2J5RNPMNHM34rICGtMrE+o7EOIll9eJa3LWhTWooerav22iKMEp3AGFxDBNdThHhrQBAJjeIZXeAuy4CV4Dz4WrWtBMXMCfxB8/gCVCY+3</latexit><latexit | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 39 | sha1_base64="xbJVLRy5zV8jtLsjAis5DXwMT4U=">AAAB7nicbVDJSgNBEK1xjXGLevTSGARPYUYEPQa9eIxgFkiG0NPpSZr0MvQihCEf4cWDIl79Hm/+jZ1kDpr4oODxXhVV9ZKMM2PD8DtYW9/Y3Nou7ZR39/YPDitHxy2jnCa0SRRXupNgQzmTtGmZ5bSTaYpFwmk7Gd/N/PYT1YYp+WgnGY0FHkqWMoKtl9o9JZh0pl+phrVwDrRKooJUoUCjX/nqDRRxgkpLODamG4WZjXOsLSOcTss9Z2iGyRgPa | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 40 | /nqDRRxgkpLODamG4WZjXOsLSOcTss9Z2iGyRgPaddTiQU1cT4/d4rOvTJAqdK+pEVz9fdEjoUxE5H4ToHtyCx7M/E/r+tsehPnTGbOUkkWi1LHkVVo9jsaME2J5RNPMNHM34rICGtMrE+o7EOIll9eJa3LWhTWooerav22iKMEp3AGFxDBNdThHhrQBAJjeIZXeAuy4CV4Dz4WrWtBMXMCfxB8/gCVCY+3</latexit> | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 41 | Unlabeled Frame B (time t+ ) <latexit | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 42 | sha1_base64="AU/vq+DAgMm1x9RhUc20aA7BJwY=">AAAB7XicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cK9gPaUDabTbt2kw27E6GE/gcvHhTx6v/x5r9x2+agrQ8GHu/NMDMvSKUw6LrfTmltfWNzq7xd2dnd2z+oHh61jco04y2mpNLdgBouRcJbKFDybqo5jQPJO8H4duZ3nrg2QiUPOEm5H9NhIiLBKFqp3Q+5RDqo1ty6OwdZJV5BalCgOah+9UPFspgnyCQ1pue5Kfo51SiY5NNKPzM8pWxMh | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 43 | ah+9UPFspgnyCQ1pue5Kfo51SiY5NNKPzM8pWxMh7xnaUJjbvx8fu2UnFklJJHSthIkc/X3RE5jYyZxYDtjiiOz7M3E/7xehtG1n4skzZAnbLEoyiRBRWavk1BozlBOLKFMC3srYSOqKUMbUMWG4C2/vEraF3XPrXv3l7XGTRFHGU7gFM7BgytowB00oQUMHuEZXuHNUc6L8+58LFpLTjFzDH/gfP4AkYyPHA==</latexit><latexit | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 44 | sha1_base64="AU/vq+DAgMm1x9RhUc20aA7BJwY=">AAAB7XicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cK9gPaUDabTbt2kw27E6GE/gcvHhTx6v/x5r9x2+agrQ8GHu/NMDMvSKUw6LrfTmltfWNzq7xd2dnd2z+oHh61jco04y2mpNLdgBouRcJbKFDybqo5jQPJO8H4duZ3nrg2QiUPOEm5H9NhIiLBKFqp3Q+5RDqo1ty6OwdZJV5BalCgOah+9UPFspgnyCQ1pue5Kfo51SiY5NNKPzM8pWxMh | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 45 | ah+9UPFspgnyCQ1pue5Kfo51SiY5NNKPzM8pWxMh7xnaUJjbvx8fu2UnFklJJHSthIkc/X3RE5jYyZxYDtjiiOz7M3E/7xehtG1n4skzZAnbLEoyiRBRWavk1BozlBOLKFMC3srYSOqKUMbUMWG4C2/vEraF3XPrXv3l7XGTRFHGU7gFM7BgytowB00oQUMHuEZXuHNUc6L8+58LFpLTjFzDH/gfP4AkYyPHA==</latexit><latexit | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 46 | sha1_base64="AU/vq+DAgMm1x9RhUc20aA7BJwY=">AAAB7XicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cK9gPaUDabTbt2kw27E6GE/gcvHhTx6v/x5r9x2+agrQ8GHu/NMDMvSKUw6LrfTmltfWNzq7xd2dnd2z+oHh61jco04y2mpNLdgBouRcJbKFDybqo5jQPJO8H4duZ3nrg2QiUPOEm5H9NhIiLBKFqp3Q+5RDqo1ty6OwdZJV5BalCgOah+9UPFspgnyCQ1pue5Kfo51SiY5NNKPzM8pWxMh | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 47 | ah+9UPFspgnyCQ1pue5Kfo51SiY5NNKPzM8pWxMh7xnaUJjbvx8fu2UnFklJJHSthIkc/X3RE5jYyZxYDtjiiOz7M3E/7xehtG1n4skzZAnbLEoyiRBRWavk1BozlBOLKFMC3srYSOqKUMbUMWG4C2/vEraF3XPrXv3l7XGTRFHGU7gFM7BgytowB00oQUMHuEZXuHNUc6L8+58LFpLTjFzDH/gfP4AkYyPHA==</latexit><latexit | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 48 | sha1_base64="AU/vq+DAgMm1x9RhUc20aA7BJwY=">AAAB7XicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cK9gPaUDabTbt2kw27E6GE/gcvHhTx6v/x5r9x2+agrQ8GHu/NMDMvSKUw6LrfTmltfWNzq7xd2dnd2z+oHh61jco04y2mpNLdgBouRcJbKFDybqo5jQPJO8H4duZ3nrg2QiUPOEm5H9NhIiLBKFqp3Q+5RDqo1ty6OwdZJV5BalCgOah+9UPFspgnyCQ1pue5Kfo51SiY5NNKPzM8pWxMh | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 49 | ah+9UPFspgnyCQ1pue5Kfo51SiY5NNKPzM8pWxMh7xnaUJjbvx8fu2UnFklJJHSthIkc/X3RE5jYyZxYDtjiiOz7M3E/7xehtG1n4skzZAnbLEoyiRBRWavk1BozlBOLKFMC3srYSOqKUMbUMWG4C2/vEraF3XPrXv3l7XGTRFHGU7gFM7BgytowB00oQUMHuEZXuHNUc6L8+58LFpLTjFzDH/gfP4AkYyPHA==</latexit>Predicted | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 50 | Pose in Frame A A Stack of Residual 3x3 Blocks Def. Conv. (dilation = 12) 2D Conv. (dilation = 12) 2D Conv. (dilation = 6) | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 51 | 2D Conv. (dilation = 24)
Def. Conv. (dilation = 6) | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 52 | Def. Conv. (dilation = 24) <latexit | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 53 | sha1_base64="N4ztu7rA6GcWPA6nADZ8tMOcvqE=">AAAB7XicbVDLSgNBEOyNrxhfUY9eBoPgKeyKoMegF48RzAOSJcxOZpMxszPLTK8QQv7BiwdFvPo/3vwbJ8keNLGgoajqprsrSqWw6PvfXmFtfWNzq7hd2tnd2z8oHx41rc4M4w2mpTbtiFouheINFCh5OzWcJpHkrWh0O/NbT9xYodUDjlMeJnSgRCwYRSc1uzqVme2VK37Vn4OskiAnFchR75W/un3NsoQrZJJa2wn8FMMJNSiY5NNSN7M8pWxEB | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 54 | 5W/un3NsoQrZJJa2wn8FMMJNSiY5NNSN7M8pWxEB7zjqKIJt+Fkfu2UnDmlT2JtXCkkc/X3xIQm1o6TyHUmFId22ZuJ/3mdDOPrcCJUmiFXbLEoziRBTWavk74wnKEcO0KZEe5WwobUUIYuoJILIVh+eZU0L6qBXw3uLyu1mzyOIpzAKZxDAFdQgzuoQwMYPMIzvMKbp70X7937WLQWvHzmGP7A+/wBz/6PRQ==</latexit><latexit | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 55 | sha1_base64="N4ztu7rA6GcWPA6nADZ8tMOcvqE=">AAAB7XicbVDLSgNBEOyNrxhfUY9eBoPgKeyKoMegF48RzAOSJcxOZpMxszPLTK8QQv7BiwdFvPo/3vwbJ8keNLGgoajqprsrSqWw6PvfXmFtfWNzq7hd2tnd2z8oHx41rc4M4w2mpTbtiFouheINFCh5OzWcJpHkrWh0O/NbT9xYodUDjlMeJnSgRCwYRSc1uzqVme2VK37Vn4OskiAnFchR75W/un3NsoQrZJJa2wn8FMMJNSiY5NNSN7M8pWxEB | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 56 | 5W/un3NsoQrZJJa2wn8FMMJNSiY5NNSN7M8pWxEB7zjqKIJt+Fkfu2UnDmlT2JtXCkkc/X3xIQm1o6TyHUmFId22ZuJ/3mdDOPrcCJUmiFXbLEoziRBTWavk74wnKEcO0KZEe5WwobUUIYuoJILIVh+eZU0L6qBXw3uLyu1mzyOIpzAKZxDAFdQgzuoQwMYPMIzvMKbp70X7937WLQWvHzmGP7A+/wBz/6PRQ==</latexit><latexit | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 57 | sha1_base64="N4ztu7rA6GcWPA6nADZ8tMOcvqE=">AAAB7XicbVDLSgNBEOyNrxhfUY9eBoPgKeyKoMegF48RzAOSJcxOZpMxszPLTK8QQv7BiwdFvPo/3vwbJ8keNLGgoajqprsrSqWw6PvfXmFtfWNzq7hd2tnd2z8oHx41rc4M4w2mpTbtiFouheINFCh5OzWcJpHkrWh0O/NbT9xYodUDjlMeJnSgRCwYRSc1uzqVme2VK37Vn4OskiAnFchR75W/un3NsoQrZJJa2wn8FMMJNSiY5NNSN7M8pWxEB | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 58 | 5W/un3NsoQrZJJa2wn8FMMJNSiY5NNSN7M8pWxEB7zjqKIJt+Fkfu2UnDmlT2JtXCkkc/X3xIQm1o6TyHUmFId22ZuJ/3mdDOPrcCJUmiFXbLEoziRBTWavk74wnKEcO0KZEe5WwobUUIYuoJILIVh+eZU0L6qBXw3uLyu1mzyOIpzAKZxDAFdQgzuoQwMYPMIzvMKbp70X7937WLQWvHzmGP7A+/wBz/6PRQ==</latexit><latexit | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 59 | sha1_base64="N4ztu7rA6GcWPA6nADZ8tMOcvqE=">AAAB7XicbVDLSgNBEOyNrxhfUY9eBoPgKeyKoMegF48RzAOSJcxOZpMxszPLTK8QQv7BiwdFvPo/3vwbJ8keNLGgoajqprsrSqWw6PvfXmFtfWNzq7hd2tnd2z8oHx41rc4M4w2mpTbtiFouheINFCh5OzWcJpHkrWh0O/NbT9xYodUDjlMeJnSgRCwYRSc1uzqVme2VK37Vn4OskiAnFchR75W/un3NsoQrZJJa2wn8FMMJNSiY5NNSN7M8pWxEB | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 60 | 5W/un3NsoQrZJJa2wn8FMMJNSiY5NNSN7M8pWxEB7zjqKIJt+Fkfu2UnDmlT2JtXCkkc/X3xIQm1o6TyHUmFId22ZuJ/3mdDOPrcCJUmiFXbLEoziRBTWavk74wnKEcO0KZEe5WwobUUIYuoJILIVh+eZU0L6qBXw3uLyu1mzyOIpzAKZxDAFdQgzuoQwMYPMIzvMKbp70X7937WLQWvHzmGP7A+/wBz/6PRQ==</latexit> | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 61 | Def. Conv. (dilation = 3) | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 62 | Def. Conv. (dilation = 18) 2D Conv. (dilation = 18)
2D Conv. (dilation = 3)
Figure 2: An illustration of our PoseWarper architecture. Given a labeled Frame A and an unlabeled
Frame B, which are separated by steps in time, our goal is to detect pose in a labeled Frame A
using the features from an unlabeled Frame B. First, we predict pose heatmaps for both frames. Then,
we compute the difference between pose heatmaps in Frame A and Frame B and feed it through a
stack of 33residual blocks. Afterwards, we attach five 33convolutional layers with dilation
ratesd2f3;6;12;18;24gand predict five sets of offsets o(d)(pn)for each pixel location pn. The
predicted offsets are used to rewarp pose heatmap B. All five rewarped heatmaps are then summed
and the resulting tensor is used to predict pose in Frame A.
to predict five sets of offsets o(d)(pn)at all pixel locations pn. The motivation for using different
dilation rates at the offset prediction stage comes from the need to consider motion cues at different | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 63 | dilation rates at the offset prediction stage comes from the need to consider motion cues at different
spatial scales. When the body motion is small, a smaller dilation rate may be more useful as it
captures subtle motion cues. Conversely, if the body motion is large, using large dilation rate allows
us to incorporate relevant information further away. Next, the predicted offsets are used to spatially
rewarp the pose heatmap fB. We do this for each of the five sets of offsets o(d), and then sum up all
five rewarped pose heatmaps to obtain a final output gA;B, which is used to predict pose in Frame A.
We implement the warping mechanism via a deformable convolution [ 28], which takes 1) the offsets
o(d)(pn), and 2) the pose heatmap fBas its inputs, and then outputs a newly sampled pose heatmap
gA;B. The subscript (A;B)is used to indicate that even though gA;Bwas resampled from tensor fB,
the offsets for rewarping were computed using A;B, which contains information from both frames.
An illustration of our architecture is presented in Figure 2.
Loss Function. As in [ 27], we use a standard pose estimation loss function which computes a mean | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 64 | An illustration of our architecture is presented in Figure 2.
Loss Function. As in [ 27], we use a standard pose estimation loss function which computes a mean
squared error between the predicted, and the ground-truth heatmaps. The ground-truth heatmap is
generated by applying a 2D Gaussian around the location of each joint.
Pose Annotation Propagation. During training, we force our model to warp pose heatmap fBfrom
an unlabeled frame B such that it would match the ground-truth pose heatmap in a labeled Frame A.
Afterwards, we can reverse the application direction of our network. This then, allows us to propagate
pose information from manually annotated frames to unlabeled frames (i.e. from a labeled Frame
A to an unlabeled Frame B). Specifically, given a pose annotation in Frame A, we can generate its
respective ground-truth heatmap yAby applying a 2D Gaussian around the location of each joint
(identically to how it was done in [ 23,27]. Then, we can predict the offsets for warping ground-truth
heatmapyAto an unlabeled Frame B, from the feature difference B;A=fB fA. Lastly, we | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 65 | heatmapyAto an unlabeled Frame B, from the feature difference B;A=fB fA. Lastly, we
use our deformable warping scheme to warp the ground-truth pose heatmap yAto Frame B, thus,
propagating pose annotations to unlabeled frames in the same video. See Figure 1 (bottom) for a
high-level illustration of this scheme.
Temporal Pose Aggregation at Inference Time. Instead of using our model to propagate pose
annotations on training videos, we can also use our deformable warping mechanism to aggregate pose
information from nearby frames during inference in order to improve the accuracy of pose detection.
For every frame at time t, we want to aggregate information from all frames at times t+where
2f 3; 2; 1;0;1;2;3g. Such a pose aggregation procedure renders pose estimation more robust
to occlusions, motion blur, and video defocus.
4
Ground Truth Reference Frame (time t)PoseWarper (time t+1)FlowNet2 (time t+1)
Figure 3: The results of a video pose propagation task by our PoseWarper and FlowNet2 [ 29]. The
first frame in each 3-frame sequence illustrates a labeled reference frame at time t. For simplicity, | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 66 | first frame in each 3-frame sequence illustrates a labeled reference frame at time t. For simplicity,
we show only the “right ankle” body joint for one person, denoted by a pink circle in each of the
frames (please zoom in for a clearer view). The second frame depicts our propagated “right ankle”
detection from the labeled frame in time t to the unlabeled frame in time t+1. The third frame shows
the propagated detection in frame t+1 produced by the FlowNet2 baseline. In contrast to our method,
FlowNet2 fails to propagate poses when there is large motion, blurriness or occlusions.
Consider a pair of frames, ItandIt+. In this case, we want to use pose information from frame
It+to improve pose detection in frame It. To do this, we first feed both frames through our trained
PoseWarper model, and obtain a spatially rewarped (resampled) pose heatmap gt;t+, which is aligned
with respect to frame Itusing the features from frame It+. We can repeat this procedure for every
value, and then aggregate pose information from multiple frames via a summation asP
gt;t+. | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 67 | value, and then aggregate pose information from multiple frames via a summation asP
gt;t+.
Implementation Details. Following the framework in [ 27], for training, we crop a 384288
bounding box around each person and use it as input to our model. During training, we use ground
truth person bounding boxes. We also employ random rotations, scaling, and horizontal flipping to
augment the data. To learn the network, we use the Adam optimizer [ 54] with a base learning rate
of10 4, which is reduced to 10 5and10 6after10, and 15epochs, respectively. The training is
performed using 4Tesla M40 GPUs, and is terminated after 20epochs. We initialize our model with
a HRNet-W48 [ 27] pretrained for a COCO keypoint estimation task. To train the deformable warping
module, we select Frame B, with a random time-gap 2[ 3;3]relative to Frame A. To compute
features relating the two frames, we use twenty 33residual blocks each with 128channels. Even
though this seems like many convolutional layers, due to a small number of channels in each layer, | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 68 | though this seems like many convolutional layers, due to a small number of channels in each layer,
this amounts to only 5:8M parameters (compared to 39M required to compute optical flow in [ 29]).
To compute the offsets o(d), we use five 33convolutional layers, each using a different dilation
rate (d= 3;6;12;18;24). To resample the pose heatmap fB, we employ five 33deformable
convolutional layers, each applied to one of the five predicted offset maps o(d). The five deformable
convolution layers too employ different dilation rates of 3;6;12;18;24. During testing, we follow
the same two-stage framework used in [ 27,23]: first, we detect the bounding boxes for each person in
the image using the detector in [ 48], and then feed the cropped images to our pose estimation model.
4 Experiments
In this section, we present our results on the PoseTrack [ 22] dataset. We demonstrate the effectiveness
of our approach on three applications: 1) video pose propagation, 2) training a network on annotations
augmented with propagated pose pseudo-labels, 3) temporal pose aggregation during inference.
5 | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 69 | augmented with propagated pose pseudo-labels, 3) temporal pose aggregation during inference.
5
Table 1: The results of video pose propagation on the PoseTrack2017 [ 22] validation set (measured in
mAP). We propagate pose information across the entire video from the manual annotations provided
in few frames. To study the effect of different levels of dilated convolutions in our PoseWarper
architecture, we also include several ablation baselines (see the bottom half of the table).
Method Head Shoulder Elbow Wrist Hip Knee Ankle Mean
Pseudo-labeling w/ HRNet [27] 79.1 86.5 81.4 74.7 81.4 79.4 72.3 79.3
Optical Flow Propagation (Farneback [55]) 76.5 82.3 74.3 69.2 80.8 74.8 70.1 75.5
Optical Flow Propagation (FlowNet2 [29]) 82.7 91.0 83.8 78.4 89.7 83.6 78.1 83.8
PoseWarper (no dilated convs) 86.1 91.7 88.0 83.5 90.2 87.3 84.6 87.2 | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 70 | PoseWarper (no dilated convs) 86.1 91.7 88.0 83.5 90.2 87.3 84.6 87.2
PoseWarper (1 dilated conv) 85.0 91.6 88.0 83.7 89.6 87.3 84.7 87.0
PoseWarper (2 dilated convs) 85.8 92.4 88.8 84.9 91.0 88.4 86.0 88.0
PoseWarper (3 dilated convs) 86.1 92.6 89.2 85.5 91.3 88.8 86.3 88.4
PoseWarper (4 dilated convs) 86.3 92.6 89.5 85.9 91.9 88.8 86.4 88.6
PoseWarper (5 dilated convs) 86.0 92.7 89.5 86.0 91.5 89.1 86.6 88.7
4.1 Video Pose Propagation
Quantitative Results. To verify that our model learns to capture pose correspondences, we apply it
to the task of video pose propagation, i.e., propagating poses across time from a few labeled frames.
Initially, we train our PoseWarper in a sparsely labeled video setting according to the procedure | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 71 | Initially, we train our PoseWarper in a sparsely labeled video setting according to the procedure
described above. In this setting, every 7thframe of a training video is labeled, i.e. there are 6
unlabeled frames between each pair of manually labeled frames. Since each video contains on
average 30 frames, we have approximately 5annotated frames uniformly spaced out in each video.
Our goal then, is to use our learned PoseWarper to propagate pose annotations from manually-labeled
frames to all unlabeled frames in the same video. Specifically, for each labeled frame in a video,
we propagate its pose information to the three preceding and three subsequent frames. We train our
PoseWarper on sparsely labeled videos from the training set of PoseTrack2017 [ 22] and then perform
our evaluations on the validation set.
To evaluate the effectiveness of our approach, we compare our model to several relevant baselines.
As our weakest baseline, we use our trained HRNet [ 27] model that simply predicts pose for every
single frame in a video. Furthermore, we also include a few propagation baselines based on warping
annotations using optical flow. The first of these uses a standard Farneback optical flow [ 55] to
warp the manually-labeled pose in each labeled frame to its three preceding and three subsequent | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 72 | warp the manually-labeled pose in each labeled frame to its three preceding and three subsequent
frames. We also include a more advanced optical flow propagation baseline that uses FlowNet2
optical flow [29]. Finally, we evaluate our PoseWarper model.
In Table 1, we present our quantitative results for video pose propagation. The evaluation is done
using an mAP metric as in [ 42]. Our best model achieves a 88:7%mAP, while the optical flow
propagation baseline using FlowNet2 [ 29] yields an accuracy of 83:8%mAP. We also note that
compared to the FlowNet2 [ 29] propagation baseline, our PoseWarper warping mechanism is not
only more accurate, but also significantly more compact ( 6M vs 39M parameters).
Ablation Studies on Dilated Convolution. In Table 1, we also present the results investigating the
effect of different levels of dilated convolutions in our PoseWarper architecture. We evaluate all these
variants on the task of video pose propagation. First, we report that removing dilated convolution
blocks from the original architecture reduces the accuracy from 88:7mAP to 87:2mAP. We also
note that a network with a single dilated convolution (using a dilation rate of 3) yields 87:0mAP. | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 73 | note that a network with a single dilated convolution (using a dilation rate of 3) yields 87:0mAP.
Adding a second dilated convolution level (using dilation rates of 3;6) improves the accuracy to 88:0.
Three dilation levels (with dilation rates of 3;6;12) yield a mAP of 88:4and four levels (dilation
rates of 3;6;12;18) give a mAP of 88:6. A network with 5dilated convolution levels yields 88:7
mAP. Adding more dilated convolutions does not improve the performance further. Additionally, we
also experimented with two networks that use dilation rates of 1;2;3;4;5, and 4;8;16;24;32, and
report that such models yield mAPs of 88:6and88:5, respectively, which are slightly lower.
Qualitative Comparison to FlowNet2. In Figure 3, we include an illustration of the motion encoded
by PoseWarper, and compare it to the optical flow computed by FlowNet2 for the video pose
propagation task. The first frame in each 3-frame sequence illustrates a labeled reference frame at
time t. For a cleaner visualization, we show only the “right ankle” body joint for one person, which is | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 74 | time t. For a cleaner visualization, we show only the “right ankle” body joint for one person, which is
marked with a pink circle in each of the frames. The second frame depicts our propagated “right
6
a) Training HRNet with Propagated Pose Pseudo Labelsb) Temporal Pose Aggregation during Inference0 100 200 300# Labeled Training Videos707580Accuracy (mAP)GT (7x)GT (1x) + pGT (6x)GT (1x)02468# Labeled Frames Per Video707580Accuracy (mAP)GT (7x)GT (1x) + pGT (6x)GT (1x)02468# Labeled Frames Per Video707580Accuracy (mAP)GT (7x)GT (1x) + T-Agg.GT (1x)0 100 200 300# Labeled Training Videos707580Accuracy (mAP)GT (7x)GT (1x) + T-Agg.GT (1x)Figure 4: A figure illustrating the value of a) training a standard HRNet [ 27] using our propagated
pose pseudo labels (left), and b) our temporal pose aggregation scheme during inference. In both
settings, we study pose detection performance as a function of 1) number of sparsely-labeled training | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 75 | settings, we study pose detection performance as a function of 1) number of sparsely-labeled training
videos (with 1manually-labeled frame per video), and 2) number of labeled frames per video (with
50sparsely-labeled videos in total). All baselines are based on retraining the standard HRNet [ 27]
model on the different training sets. The "GT (1x)" baseline is trained in a standard way on sparsely
labeled video data. The "GT (7x)" baseline uses 7x more manually annotated data relative to the
"GT (1x)" baseline. Our approach on the left subfigure ("GT (1x) + pGT (6x)"), augments the
original sparsely labeled video data with our propagated pose pseudo labels ( 6nearby frames for
every manually-labeled frame). Lastly, in b) "GT (1x) + T-Agg" denotes the use of PoseWarper to
fuse pose information from multiple neighboring frames during inference (training is done as in "GT
(1x)" baseline). From the results, we observe that both application modalities of PoseWarper provide
an effective way to achieve strong pose accuracy while reducing the number of manual annotations.
ankle” detection from the labeled frame in time t to the unlabeled frame in time t+1. The third frame | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 76 | ankle” detection from the labeled frame in time t to the unlabeled frame in time t+1. The third frame
shows the propagated detection in frame t+1 produced by the FlowNet2 baseline. These results
suggest that FlowNet2 struggles to accurately warp poses if 1) there is large motion, 2) occlusions, or
3) blurriness. In contrast, our PoseWarper handles these cases robustly, which is also indicated by our
results in Table 1 (i.e., 88:7vs83:8mAP w.r.t. FlowNet2).
4.2 Data Augmentation with PoseWarper
Here we consider the task of propagating poses on sparsely labeled training videos using PoseWarper,
and then using them as pseudo-ground truth labels (in addition to the original manual labels) to
train a standard HRNet-W48 [ 27]. For this experiment, we study the pose detection accuracy as
a function of two variables: 1) the total number of sparsely-labeled videos, and 2) the number of
manually-annotated frames per video. We aim to study how much we can reduce manual labeling
through our mechanism of pose propagation, while maintaining strong pose accuracy. Note, that we
first train our PoseWarper on sparsely labeled videos from the training set of PoseTrack2017 [ 22]. | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 77 | first train our PoseWarper on sparsely labeled videos from the training set of PoseTrack2017 [ 22].
Then, we propagate pose annotations on the same set of training videos. Afterwards, we retrain the
model on the joint training set comprised of sparse manual pose annotations and our propagated
poses. Lastly, we evaluate this trained model on the validation set.
All results are based on a standard HRNet [ 27] model trained on different forms of training data. "GT
(1x)" refers to a model trained on sparsely labeled videos using ground-truth annotations only. "GT
(7x)" baseline employs 7x more manually-annotated poses relative to "GT (1x)" (the annotations
are part of the PoseTrack2017 training set). In comparison, our approach ("GT (1x) + pGT (6x)"),
is trained on a joint training set consisting of sparse manual pose annotations (same as "GT (1x)"
baseline) and our propagated poses (on the training set of PoseTrack2017), which we use as pseudo
ground truth data (pGT). As before, for every labeled frame we propagate the ground truth pose to
the3previous and the 3subsequent frames, which allows us to expand the training set by 7times. | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 78 | the3previous and the 3subsequent frames, which allows us to expand the training set by 7times.
Based on the results in the left subfigure of Figure 4, we can draw several conclusions. First, we
note that when there are very few labeled videos (i.e., 5), all three baselines perform poorly (leftmost
figure). This indicates that in this setting there is not enough data to learn an effective pose detection
model. Second, we observe that when the number of labeled videos is somewhat reasonable (e.g.,
50 100), our approach significantly outperforms the "GT (1x)" baseline, and is only slightly worse
relative to the "GT (7x)" baseline. As we increase the number of labeled videos, the gaps among the
three methods shrink, suggesting that the model becomes saturated.
7
Table 2: Multi-person pose estimation results on the validation and test sets of PoseTrack2017
and PoseTrack2018 datasets. Even though our model is designed to improve pose detection in
scenarios involving sparsely-labeled videos, here we show that our temporal pose aggregation scheme
during inference is also useful for models trained on densely labeled videos. We improve upon the
state-of-the-art single-frame baselines [23, 27, 56]. | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 79 | during inference is also useful for models trained on densely labeled videos. We improve upon the
state-of-the-art single-frame baselines [23, 27, 56].
Dataset Method Head Shoulder Elbow Wrist Hip Knee Ankle Mean
PoseTrack17 Val SetGirdhar et al. [48] 72.8 75.6 65.3 54.3 63.5 60.9 51.8 64.1
Xiu et al. [57] 66.7 73.3 68.3 61.1 67.5 67.0 61.3 66.5
Bin et al [23] 81.7 83.4 80.0 72.4 75.3 74.8 67.1 76.7
HRNet [27] 82.1 83.6 80.4 73.3 75.5 75.3 68.5 77.3
MDPN [56] 85.2 88.5 83.9 77.5 79.0 77.0 71.4 80.7
PoseWarper 81.4 88.3 83.9 78.0 82.4 80.5 73.6 81.2
PoseTrack17 Test SetGirdhar et al. [48] - - - - - - - 59.6 | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 80 | PoseTrack17 Test SetGirdhar et al. [48] - - - - - - - 59.6
Xiu et al. [57] 64.9 67.5 65.0 59.0 62.5 62.8 57.9 63.0
Bin et al [23] 80.1 80.2 76.9 71.5 72.5 72.4 65.7 74.6
HRNet [27] 80.1 80.2 76.9 72.0 73.4 72.5 67.0 74.9
PoseWarper 79.5 84.3 80.1 75.8 77.6 76.8 70.8 77.9
PoseTrack18 Val SetAlphaPose [58] 63.9 78.7 77.4 71.0 73.7 73.0 69.7 71.9
MDPN [56] 75.4 81.2 79.0 74.1 72.4 73.0 69.9 75.0
PoseWarper 79.9 86.3 82.4 77.5 79.8 78.8 73.2 79.7
PoseTrack18 Test SetAlphaPose++ [56, 58] - - - 66.2 - - 65.0 67.6
MDPN [56] - - - 74.5 - - 69.0 76.4 | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 81 | MDPN [56] - - - 74.5 - - 69.0 76.4
PoseWarper 78.9 84.4 80.9 76.8 75.6 77.5 71.8 78.0
As we vary the number of labeled frames per video (second leftmost figure), we notice several
interesting patterns. First, we note that for a small number of labeled frames per video (i.e., 1 2) our
approach outperforms the "GT (1x)" baseline by a large margin. Second, we note that the performance
of our approach and the "GT (7x)" becomes very similar as we add 2or more labeled frames per
video. These findings further strengthen our previous observation that PoseWarper allows us to reduce
the annotation cost without a significant loss in performance.
4.3 Improved Pose Estimation via Temporal Pose Aggregation
In this subsection we assess the ability of PoseWarper to improve the accuracy of pose estimation at
test time by using our deformable warping mechanism to aggregate pose information from nearby
frames. We visualize our results in Figure 4 b), where we evaluate the effectiveness of our temporal
pose aggregation during inference for models trained a) with a different number of labeled videos | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 82 | frames. We visualize our results in Figure 4 b), where we evaluate the effectiveness of our temporal
pose aggregation during inference for models trained a) with a different number of labeled videos
(second rightmost figure), and b) with a different number of manually-labeled frames per video
(rightmost figure). We compare our approach ("GT (1x) + T-Agg.") to the same "GT (7x)" and "GT
(1x)" baselines defined in the previous subsection. Note that our method in this case is trained exactly
as "GT (1x)" baseline, the only difference comes from the inference procedure.
When the number of training videos and/or manually labeled frames is small, our approach provides
a significant accuracy boost with respect to the "GT (1x)" baseline. However, once, we increase the
number of labeled videos/frames, the gap between all three baselines shrinks, and the model becomes
more saturated. Thus, our temporal pose aggregation scheme during inference is another effective
way to maintain strong performance in a sparsely-labeled video setting.
4.4 Comparison to State-of-the-Art
We also test the effectiveness of our temporal pose aggregation scheme, when the model is trained on | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 83 | 4.4 Comparison to State-of-the-Art
We also test the effectiveness of our temporal pose aggregation scheme, when the model is trained on
the full PoseTrack [ 22] dataset. Table 2 compares our method to the most recent approaches in this
area [ 48,57,23,27]. These results suggest that although we designed our method to improve pose
estimation when training videos are sparsely-labeled, our temporal pose aggregation scheme applied
at inference is also useful for models trained on densely-labeled videos. Our PoseWarper obtains
81:2mAP and 77:9mAP on PoseTrack2017 validation and test sets respectively, and 79:7mAP and
78:0mAP on PoseTrack2018 validation and test sets respectively, thus outperforming prior single
frame baselines [48, 57, 23, 27].
8
Frame tFrame t+5Farneback FlowPredicted Human MotionOffset MagnitudesChannel 99 (x,y)Channel 123 (x,y)
Figure 5: In the first two columns, we show a pair of video frames used as input for our model.
The3rdand4thcolumns depict 2randomly selected offset channels visualized as a motion field.
Different channels appear to capture the motion of different body parts. In the 5thcolumn, we display | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 84 | Different channels appear to capture the motion of different body parts. In the 5thcolumn, we display
the offset magnitudes, which highlight salient human motion. Finally, the last two columns illustrate
the standard Farneback flow, and the human motion predicted from our learned offsets. To predict
human motion we train a linear classifier to regress the ground-truth (x;y)displacement of each
joint from the offset maps. The color wheel, at the bottom right corner encodes motion direction.
4.5 Interpreting Learned Offsets
Understanding what information is encoded in our learned offsets is nearly as difficult as analyzing
any other CNN features [ 59,60]. The main challenge comes from the high dimensionality of offsets:
we are predicting ckhkw(x;y)displacements for every pixel for each of the five dilation
ratesd, wherecis the number of channels, and kh;kware the convolutional kernel height and width
respectively.
In columns 3;4of Figure 5, we visualize two randomly-selected offset channels as a motion field.
Based on this figure, it appears that different offset maps encode different motions rather than all | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 85 | Based on this figure, it appears that different offset maps encode different motions rather than all
predicting the same solution (say, the optical flow between the two frames). This makes sense, as the
network may decide to ignore motions of uninformative regions, and instead capture the motion of
different human body parts in different offset maps (say, a hand as opposed to the head). We also note
that the magnitudes of our learned offsets encode salient human motion (see Column 5 of Figure 5).
Lastly, to verify that our learned offsets encode human motion, for each point pndenoting a body joint,
we extract our predicted offsets and train a linear classifier to regress the ground truth (x;y)motion
displacement of this body joint. In Column 7 of Figure 5, we visualize our predicted motion outputs
for every pixel. We show Farneback’s optical flow in Column 6. Note that in regions containing
people, our predicted human motion matches Farneback optical flow. Furthermore, we point out that
compared to the standard Farneback optical flow, our motion fields look less noisy.
5 Conclusions
In this work, we introduced PoseWarper, a novel architecture for pose detection in sparsely labeled | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 86 | 5 Conclusions
In this work, we introduced PoseWarper, a novel architecture for pose detection in sparsely labeled
videos. Our PoseWarper can be effectively used for multiple applications, including video pose
propagation, and temporal pose aggregation. In these settings, we demonstrated that our approach
reduces the need for densely labeled video data, while producing strong pose detection performance.
Furthermore, our state-of-the-art results on PoseTrack2017 and PoseTrack2018 datasets demonstrate
that our PoseWarper is useful even when the training videos are densely-labeled. Our future work
involves improving our model ability to propagate labels and aggregate temporal information when
the input frames are far away from each other. We are also interested in exploring self-supervised
learning objectives for our task, which may further reduce the need of pose annotations in video. We
will release our source code and our trained models upon publication of the paper.
References
[1]Gedas Bertasius, Jianbo Shi, and Lorenzo Torresani. Deepedge: A multi-scale bifurcated deep network
for top-down contour detection. In The IEEE Conference on Computer Vision and Pattern Recognition
(CVPR) , June 2015.
9 | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 87 | for top-down contour detection. In The IEEE Conference on Computer Vision and Pattern Recognition
(CVPR) , June 2015.
9
[2]Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In CVPR , 2017.
[3]Alexander Toshev and Christian Szegedy. Deeppose: Human pose estimation via deep neural networks. In
CVPR , 2014.
[4]Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, and Jianbo Shi. Convolutional random walk networks
for semantic image segmentation. In The IEEE Conference on Computer Vision and Pattern Recognition
(CVPR) , July 2017.
[5]Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Spatial pyramid pooling in deep convolutional
networks for visual recognition. In Computer Vision – ECCV 2014 , 2014.
[6]Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask R-CNN. In Proceedings of the
International Conference on Computer Vision (ICCV) , 2017. | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 88 | International Conference on Computer Vision (ICCV) , 2017.
[7]Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollár. Focal Loss for Dense Object
Detection. In Proceedings of the International Conference on Computer Vision (ICCV) , 2017.
[8]Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards real-time object
detection with region proposal networks. In Neural Information Processing Systems (NIPS) , 2015.
[9]Ross Girshick. Fast R-CNN. In Proceedings of the International Conference on Computer Vision (ICCV) ,
2015.
[10] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate
object detection and semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision
and Pattern Recognition (CVPR) , 2014.
[11] Saurabh Gupta, Ross Girshick, Pablo Arbelaez, and Jitendra Malik. Learning rich features from RGB-D
images for object detection and segmentation. In ECCV , 2014.
[12] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 89 | [12] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and
Alexander C. Berg. Ssd: Single shot multibox detector. In ECCV , 2016.
[13] Jifeng Dai, Yi Li, Kaiming He, and Jian Sun. R-fcn: Object detection via region-based fully convolutional
networks. In Advances in Neural Information Processing Systems 29 , pages 379–387. Curran Associates,
Inc., 2016.
[14] Joseph Redmon, Santosh Kumar Divvala, Ross B. Girshick, and Ali Farhadi. You only look once: Unified,
real-time object detection. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR
2016, Las Vegas, NV , USA, June 27-30, 2016 , pages 779–788, 2016.
[15] Joseph Redmon and Ali Farhadi. YOLO9000: better, faster, stronger. In 2017 IEEE Conference on
Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017 , pages
6517–6525, 2017. | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 90 | Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017 , pages
6517–6525, 2017.
[16] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. Imagenet classification with deep convolutional
neural networks. In NIPS , 2012.
[17] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru
Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Computer Vision
and Pattern Recognition (CVPR) , 2015.
[18] K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recognition. In
ICLR , 2015.
[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition.
2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages 770–778, 2016.
[20] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. ImageNet: A Large-Scale Hierarchical Image | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 91 | Database. In CVPR09 , 2009.
[21] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár,
and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In European Conference on
Computer Vision (ECCV) , Zürich, September 2014.
[22] Umar Iqbal, Anton Milan, and Juergen Gall. Posetrack: Joint multi-person pose estimation and tracking.
InIEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2017.
10
[23] Bin Xiao, Haiping Wu, and Yichen Wei. Simple baselines for human pose estimation and tracking. In
European Conference on Computer Vision (ECCV) , 2018.
[24] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using
part affinity fields. In CVPR , 2017.
[25] Alejandro Newell, Kaiyu Yang, and Jia Deng. Stacked hourglass networks for human pose estimation. In
Computer Vision - ECCV 2016 - 14th European Conference, Amsterdam, The Netherlands, October 11-14, | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 92 | Computer Vision - ECCV 2016 - 14th European Conference, Amsterdam, The Netherlands, October 11-14,
2016, Proceedings, Part VIII , pages 483–499, 2016.
[26] Shih-En Wei, Varun Ramakrishna, Takeo Kanade, and Yaser Sheikh. Convolutional pose machines. In
CVPR , pages 4724–4732. IEEE Computer Society, 2016.
[27] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human
pose estimation. In CVPR , 2019.
[28] J. Dai, H. Qi, Y . Xiong, Y . Li, G. Zhang, H. Hu, and Y . Wei. Deformable convolutional networks. In 2017
IEEE International Conference on Computer Vision (ICCV) , volume 00, pages 764–773, Oct. 2017.
[29] Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox.
Flownet 2.0: Evolution of optical flow estimation with deep networks. CoRR , abs/1612.01925, 2016.
[30] Mykhaylo Andriluka, Stefan Roth, and Bernt Schiele. Pictorial structures revisited: People detection and | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 93 | [30] Mykhaylo Andriluka, Stefan Roth, and Bernt Schiele. Pictorial structures revisited: People detection and
articulated pose estimation. In CVPR , pages 1014–1021. IEEE Computer Society, 2009.
[31] M. Andriluka, S. Roth, and B. Schiele. Monocular 3d pose estimation and tracking by detection. In 2010
IEEE Computer Society Conference on Computer Vision and Pattern Recognition , pages 623–630, June
2010.
[32] Sam Johnson and Mark Everingham. Clustered pose and nonlinear appearance models for human pose
estimation. In Proc. BMVC , pages 12.1–11, 2010. doi:10.5244/C.24.12.
[33] Leonid Pishchulin, Mykhaylo Andriluka, Peter V . Gehler, and Bernt Schiele. Strong appearance and
expressive spatial models for human pose estimation. In ICCV , pages 3487–3494. IEEE Computer Society,
2013.
[34] Yi Yang and D. Ramanan. Articulated pose estimation with flexible mixtures-of-parts. In Proceedings of
the 2011 IEEE Conference on Computer Vision and Pattern Recognition , CVPR ’11, pages 1385–1392, | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 94 | the 2011 IEEE Conference on Computer Vision and Pattern Recognition , CVPR ’11, pages 1385–1392,
Washington, DC, USA, 2011. IEEE Computer Society.
[35] M. Dantone, J. Gall, C. Leistner, and L. van Gool. Human pose estimation using body parts dependent
joint regressors. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR) , pages 3041–3048,
Portland, OR, USA, June 2013. IEEE.
[36] Xiangyang Lan and D. P. Huttenlocher. Beyond trees: common-factor models for 2d human pose recovery.
InTenth IEEE International Conference on Computer Vision (ICCV’05) Volume 1 , volume 1, pages
470–477 V ol. 1, Oct 2005.
[37] L. Sigal and M. J. Black. Measure locally, reason globally: Occlusion-sensitive articulated pose estimation.
InProc. IEEE Conf. on Computer Vision and Pattern Recognition, CVPR , volume 2, pages 2041–2048,
New York, NY , June 2006.
[38] Yang Wang and Greg Mori. Multiple tree models for occlusion and spatial constraints in human pose
estimation. In David Forsyth, Philip Torr, and Andrew Zisserman, editors, Computer Vision – ECCV 2008 , | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 95 | estimation. In David Forsyth, Philip Torr, and Andrew Zisserman, editors, Computer Vision – ECCV 2008 ,
pages 710–724, Berlin, Heidelberg, 2008. Springer Berlin Heidelberg.
[39] J. Carreira, P. Agrawal, K. Fragkiadaki, and J. Malik. Human pose estimation with iterative error feedback.
In2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages 4733–4742, June
2016.
[40] Xianjie Chen and Alan L Yuille. Articulated pose estimation by a graphical model with image dependent
pairwise relations. In Z. Ghahramani, M. Welling, C. Cortes, N. D. Lawrence, and K. Q. Weinberger,
editors, Advances in Neural Information Processing Systems 27 , pages 1736–1744. Curran Associates,
Inc., 2014.
[41] W. Ouyang, X. Chu, and X. Wang. Multi-source deep learning for human pose estimation. In 2014 IEEE
Conference on Computer Vision and Pattern Recognition , pages 2337–2344, June 2014.
11 | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 96 | Conference on Computer Vision and Pattern Recognition , pages 2337–2344, June 2014.
11
[42] Eldar Insafutdinov, Leonid Pishchulin, Bjoern Andres, Mykhaylo Andriluka, and Bernt Schieke. Deepercut:
A deeper, stronger, and faster multi-person pose estimation model. In European Conference on Computer
Vision (ECCV) , 2016.
[43] Leonid Pishchulin, Eldar Insafutdinov, Siyu Tang, Bjoern Andres, Mykhaylo Andriluka, Peter Gehler, and
Bernt Schiele. Deepcut: Joint subset partition and labeling for multi person pose estimation. In IEEE
Conference on Computer Vision and Pattern Recognition (CVPR) , 2016.
[44] Jonathan J Tompson, Arjun Jain, Yann LeCun, and Christoph Bregler. Joint training of a convolutional
network and a graphical model for human pose estimation. In Z. Ghahramani, M. Welling, C. Cortes, N. D.
Lawrence, and K. Q. Weinberger, editors, Advances in Neural Information Processing Systems 27 , pages
1799–1807. Curran Associates, Inc., 2014. | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 97 | Lawrence, and K. Q. Weinberger, editors, Advances in Neural Information Processing Systems 27 , pages
1799–1807. Curran Associates, Inc., 2014.
[45] George Papandreou, Tyler Zhu, Nori Kanazawa, Alexander Toshev, Jonathan Tompson, Chris Bregler, and
Kevin Murphy. Towards accurate multi-person pose estimation in the wild. In CVPR , pages 3711–3719.
IEEE Computer Society, 2017.
[46] Vasileios Belagiannis and Andrew Zisserman. Recurrent human pose estimation. In International
Conference on Automatic Face and Gesture Recognition . IEEE, 2017.
[47] Eldar Insafutdinov, Mykhaylo Andriluka, Leonid Pishchulin, Siyu Tang, Evgeny Levinkov, Bjoern Andres,
and Bernt Schiele. Articulated multi-person tracking in the wild. In 2017 IEEE Conference on Computer
Vision and Pattern Recognition (CVPR) , pages 1293–1301. IEEE, July 2017. Oral.
[48] Rohit Girdhar, Georgia Gkioxari, Lorenzo Torresani, Manohar Paluri, and Du Tran. Detect-and-Track:
Efficient Pose Estimation in Videos. In CVPR , 2018. | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 98 | Efficient Pose Estimation in Videos. In CVPR , 2018.
[49] Jie Song, Limin Wang, Luc Van Gool, and Otmar Hilliges. Thin-slicing network: A deep structured model
for pose estimation in videos. In CVPR , pages 4420–4229, 2017.
[50] J. Charles, T. Pfister, D. Magee, D. Hogg, and A. Zisserman. Personalizing human video pose estimation.
InIEEE Conference on Computer Vision and Pattern Recognition , 2016.
[51] T. Pfister, J. Charles, and A. Zisserman. Flowing convnets for human pose estimation in videos. In IEEE
International Conference on Computer Vision , 2015.
[52] Dingwen Zhang, Guangyu Guo, Dong Huang, and Junwei Han. Poseflow: A deep motion representation
for understanding human behaviors in videos. In The IEEE Conference on Computer Vision and Pattern
Recognition (CVPR) , June 2018.
[53] O. Wiles, A.S. Koepke, and A. Zisserman. Self-supervised learning of a facial attribute embedding from
video. In British Machine Vision Conference , 2018. | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
1906.04016 | 99 | video. In British Machine Vision Conference , 2018.
[54] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR , 2015.
[55] Gunnar Farnebäck. Two-frame motion estimation based on polynomial expansion. In Proceedings of the
13th Scandinavian Conference on Image Analysis , SCIA’03, pages 363–370, Berlin, Heidelberg, 2003.
Springer-Verlag.
[56] Hengkai Guo, Tang Tang, Guozhong Luo, Riwei Chen, Yongchen Lu, and Linfu Wen. Multi-domain pose
network for multi-person pose estimation and tracking. In Computer Vision - ECCV 2018 Workshops Munich, Germany, September 8-14, 2018, Proceedings, Part II , pages 209–216, 2018.
[57] Yuliang Xiu, Jiefeng Li, Haoyu Wang, Yinghong Fang, and Cewu Lu. Pose Flow: Efficient online pose
tracking. In BMVC , 2018.
[58] Hao-Shu Fang, Shuqin Xie, Yu-Wing Tai, and Cewu Lu. Rmpe: Regional multi-person pose estimation. In
The IEEE International Conference on Computer Vision (ICCV) , Oct 2017. | 1906.04016 | Learning Temporal Pose Estimation from Sparsely-Labeled Videos | Modern approaches for multi-person pose estimation in video require large
amounts of dense annotations. However, labeling every frame in a video is
costly and labor intensive. To reduce the need for dense annotations, we
propose a PoseWarper network that leverages training videos with sparse
annotations (every k frames) to learn to perform dense temporal pose
propagation and estimation. Given a pair of video frames---a labeled Frame A
and an unlabeled Frame B---we train our model to predict human pose in Frame A
using the features from Frame B by means of deformable convolutions to
implicitly learn the pose warping between A and B. We demonstrate that we can
leverage our trained PoseWarper for several applications. First, at inference
time we can reverse the application direction of our network in order to
propagate pose information from manually annotated frames to unlabeled frames.
This makes it possible to generate pose annotations for the entire video given
only a few manually-labeled frames. Compared to modern label propagation
methods based on optical flow, our warping mechanism is much more compact (6M
vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also
show that we can improve the accuracy of a pose estimator by training it on an
augmented dataset obtained by adding our propagated poses to the original
manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose
information from neighboring frames during inference. This allows our system to
achieve state-of-the-art pose detection results on the PoseTrack2017 and
PoseTrack2018 datasets. Code has been made available at:
https://github.com/facebookresearch/PoseWarper. | http://arxiv.org/pdf/1906.04016 | [
"Gedas Bertasius",
"Christoph Feichtenhofer",
"Du Tran",
"Jianbo Shi",
"Lorenzo Torresani"
] | [
"cs.CV"
] | Accepted to NeurIPS 2019 | null | cs.CV | 20190606 | 20191211 | [
{
"id": "1412.1123",
"title": "1. DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection",
"authors": "Gedas Bertasius, Jianbo Shi, Lorenzo Torresani",
"year": "2015"
},
{
"id": "1611.05431",
"title": "2. Aggregated Residual Transformations for Deep Neural Networks",
"authors": "Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He",
"year": "2017"
},
{
"id": "1312.4659",
"title": "3. DeepPose: Human Pose Estimation via Deep Neural Networks",
"authors": "Alexander Toshev, Christian Szegedy",
"year": "2014"
},
{
"id": "1605.07681",
"title": "4. Convolutional Random Walk Networks for Semantic Image Segmentation",
"authors": "Gedas Bertasius, Lorenzo Torresani, Stella X. Yu, Jianbo Shi",
"year": "2017"
},
{
"id": "1406.4729",
"title": "5. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"authors": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun",
"year": "2014"
},
{
"id": "1703.06870",
"title": "6. Mask R-CNN",
"authors": "Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick",
"year": "2018"
}
] |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.