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 | 100 | The IEEE International Conference on Computer Vision (ICCV) , Oct 2017.
[59] Matthew D. Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. CoRR ,
abs/1311.2901, 2013.
[60] Jason Yosinski, Jeff Clune, Anh Mai Nguyen, Thomas J. Fuchs, and Hod Lipson. Understanding neural
networks through deep visualization. CoRR , abs/1506.06579, 2015.
12 | 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"
}
] |
1912.04573 | 0 | Classifying, Segmenting, and Tracking
Object Instances in Video with Mask Propagation
Gedas Bertasius, Lorenzo Torresani
Facebook AI
Abstract
We introduce a method for simultaneously classifying,
segmenting and tracking object instances in a video sequence. Our method, named MaskProp, adapts the popular Mask R-CNN to video by adding a mask propagation
branch that propagates frame-level object instance masks
from each video frame to all the other frames in a video clip.
This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip. Clip-level instance tracks generated
densely for each frame in the sequence are finally aggregated to produce video-level object instance segmentation
and classification. Our experiments demonstrate that our
clip-level instance segmentation makes our approach robust
to motion blur and object occlusions in video. MaskProp
achieves the best reported accuracy on the YouTube-VIS
dataset, outperforming the ICCV 2019 video instance segmentation challenge winner despite being much simpler and
using orders of magnitude less labeled data (1.3M vs 1B images and 860K vs 14M bounding boxes).
1. Introduction | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 1 | using orders of magnitude less labeled data (1.3M vs 1B images and 860K vs 14M bounding boxes).
1. Introduction
In this paper, we tackle the recently introduced video instance segmentation problem [42]. This task requires segmenting all instances of a predefined set of object classes
in each frame, classifying them, and linking individual instances over the entire sequence.
In recent years, convolutional networks have obtained remarkable results in still-image object detection [16, 33, 14,
15], and segmentation [27, 45, 8, 4]. However, extending
these models to video instance segmentation is challenging. In order to localize objects precisely, these methods
have to operate at very large spatial resolution. As a result,
detectors based on the popular ResNet-101 or ResNet-152
backbones [17] can rarely fit more than one image per GPU
during training. In the context of video instance segmentation this is problematic because tracking objects over time
requires analyzing multiple video frames simultaneously.
To address this issue, one could reduce the spatial resFigure 1: In this paper, we tackle the problem of video instance segmentation, which requires classifying, segmenting, and tracking object instances in a given video sequence.
Our proposed Mask Propagation framework (MaskProp)
provides a simple and effective way for solving this task. | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 2 | Our proposed Mask Propagation framework (MaskProp)
provides a simple and effective way for solving this task.
olution of the input and fit more video frames in a GPU.
However, doing so typically leads to a significant drop in
segmentation or detection performance. Alternatively, one
could perform high-resolution instance segmentation on individual frames and then link segmentations temporally in
a separate post-processing stage. However, performing instance segmentation and tracking in two disjoint steps often produces suboptimal results, because these two tasks
are closely intertwined. The key challenge then becomes
designing a unified model that can track objects in video
while maintaining strong detection accuracy.
Currently, the best method for video instance segmentation is the ICCV 2019 challenge winner [28]. It tackles
video instance segmentation by dividing it into four problems: 1) detection, 2) classification, 3) segmentation, and
4) tracking. These four problems are solved independently
using several off-the-shelf components and their respective
solutions are combined and adapted to the video instance
segmentation task. However, despite effective performance,
such an approach is disadvantageous because it requires designing and tuning a separate model (or, in some cases, an | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 3 | segmentation task. However, despite effective performance,
such an approach is disadvantageous because it requires designing and tuning a separate model (or, in some cases, an
ensemble of models) for each of the four tasks. This renders the approach costly and cumbersome. On the other
end of the complexity spectrum, MaskTrack R-CNN [42] is
a simple unified approach trained end-to-end but it achieves
significantly lower performance ( 30:3vs44:8video mAP).
To address the shortcomings of these prior methods wearXiv:1912.04573v4 [cs.CV] 9 Jul 2021
MaskTrack R-CNN [42] ICCV19 Challenge Winner [28] MaskProp
ModelClassification cls head Mask R-CNN [16], ResNeXt-101 32x48d [30] cls head
Localization bbox head Mask R-CNN [16] bbox head
Segmentation mask head DeepLabv3 [9], Box2Seg [29] mask head
Tracking tracking head UnOVOST [47], ReID Net [18, 31] mask propagation head
Optical Flow - PWC-Net [35] Pre-training DatasetsImageNet [34] (1.3M images) X X X | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 4 | Optical Flow - PWC-Net [35] Pre-training DatasetsImageNet [34] (1.3M images) X X X
COCO [25] (860K bboxes) X X X
Instagram [30] (1B images) - X OpenImages [23] (14M bboxes) - X Performancevideo mAP 30.3 44.8 46.6
video AP@75 32.6 48.9 51.2
Table 1: A table comparing our work to prior video instance segmentation methods [42, 28]. The ICCV 2019 Challenge
Winner [28] decomposes video instance segmentation into four different problems, solves each of them independently using
ensembles of different models, and then combines these solutions. In contrast, our approach relies on a single unified model
trained end-to-end. Despite being simpler, and using several orders of magnitude less pretraining data (1.3M vs 1B images
and 860K vs 14M bounding boxes) our model achieves higher accuracy. Furthermore, compared to MaskTrack R-CNN [42],
our work yields a 16.3% gain in mAP (46.6% vs 30.3%).
introduce MaskProp, a simple mask propagation framework
for simultaneously classifying, segmenting and tracking object instances in video. Our method adapts the popular | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 5 | introduce MaskProp, a simple mask propagation framework
for simultaneously classifying, segmenting and tracking object instances in video. Our method adapts the popular
Mask R-CNN [16] to video by adding a branch that propagates frame-level instance masks from each video frame
to other frames within a temporal neighborhood (which we
refer to as a clip). This allows our method to compute cliplevel instance tracks centered at each individual frame of the
video. These densely estimated clip-level tracks are then
aggregated to form accurate and coherent object instance
sequences for the entire video, regardless of its length. This
renders our approach capable of handling challenging cases
of occlusions, disocclusions, and motion blur. Our method
achieves the best reported accuracy on the YouTube-VIS
dataset [42], outperforming the ICCV 2019 challenge winner [28] despite being much simpler and using significantly
less labeled data (1000x fewer images and 10x fewer bounding boxes). In Table 1, we compare our approach vs these
prior methods in terms of accuracy and other characteristics.
2. Related Work
Instance Segmentation in Images. Compared to instance
segmentation in images [11, 16, 44, 1, 2, 22, 26], the problem considered in this paper requires not only to segment
object instances in individual frames, but also to determine | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 6 | object instances in individual frames, but also to determine
instance correspondences across multiple frames. We leverage the Mask R-CNN model [16] for still-image instance
segmentation and adapt it to track object instances in video.
Object Detection in Video. Object detection in video requires classifying and localizing objects in every frame of
a given video. Most modern video object detection systems [46, 5, 40, 13] implement some form of spatiotemporal feature alignment for improving object detection accuracy in individual video frames. However, these systems
are typically not designed for tracking object instances. In
contrast, our mask propagation produces clip-level instancesegmentations rather than frame-level bounding boxes.
Video Object Segmentation. The task of video object
segmentation requires segmenting foreground objects in a
class-agnostic fashion [43, 21, 36, 38], often by leveraging
ground truth masks available for the first frame during inference [6, 32, 10, 19, 37]. Instead, video instance segmentation requires finding all instances of a predefined set of
object classes in each frame, classifying them and linking
them over the entire sequence.
Video Instance Segmentation. The recently introduced
video instance segmentation task [42] requires classifying,
segmenting and tracking object instances in videos. This | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 7 | Video Instance Segmentation. The recently introduced
video instance segmentation task [42] requires classifying,
segmenting and tracking object instances in videos. This
is the task considered in this work. There are only a few
video instance segmentation methods we can compare our
approach to. The MaskTrack R-CNN [42] presents a unified model for video instance segmentation. It augments the
original Mask R-CNN [16] with a tracking branch that establishes associations among object instances segmented in
separate frames. Furthermore, we include the ICCV 2019
video instance segmentation challenge winner [28] in our
comparison. This approach divides video instance segmentation into four separate subproblems: classification, detection, segmentation, and tracking. A separate model (or
an ensemble of models) is used to solve each of these subproblems, and these solutions are then combined to produce
video instance segmentation results. For brevity, from now
on we refer to it as EnsembleVIS to indicate that it is an ensemble approach designed for video instance segmentation.
Our MaskProp framework provides advantages over
both of these methods [42, 28]. Similarly to MaskTrack
R-CNN [42], our method is a unified and simple approach. | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 8 | both of these methods [42, 28]. Similarly to MaskTrack
R-CNN [42], our method is a unified and simple approach.
However, our mask propagation branch is much more effective than the tracking branch of MaskTrack R-CNN, achieving much higher accuracy relative to this baseline. Furthermore, compared to EnsembleVIS [28], our method 1)
is much simpler, 2) uses significantly less labeled data, and | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 9 | clip-level instance tracktemporally propagated instance featuresfeature tensorinstance feature computation at time t frame-level instance mask time ttime t⌦<latexit | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 10 | sha1_base64="f9ZxFFopka18yU5lyTwVt2fcbbg=">AAAB7nicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Bj04jGCeUCyhNnJbDJkdmaZ6RXCko/w4kERr36PN//GSbIHTSxoKKq66e6KUiks+v63t7a+sbm1Xdop7+7tHxxWjo5bVmeG8SbTUptORC2XQvEmCpS8kxpOk0jydjS+m/ntJ26s0OoRJykPEzpUIhaMopPaPY0i4bZfqfo1fw6ySoKCVKFAo1/56g00yxKukElqbTfwUwxzalAwyaflXmZ5StmYD | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 11 | 1/56g00yxKukElqbTfwUwxzalAwyaflXmZ5StmYDnnXUUXdkjCfnzsl504ZkFgbVwrJXP09kdPE2kkSuc6E4sguezPxP6+bYXwT5kKlGXLFFoviTBLUZPY7GQjDGcqJI5QZ4W4lbEQNZegSKrsQguWXV0nrshb4teDhqlq/LeIowSmcwQUEcA11uIcGNIHBGJ7hFd681Hvx3r2PReuaV8ycwB94nz+F64+t</latexit><latexit | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 12 | sha1_base64="f9ZxFFopka18yU5lyTwVt2fcbbg=">AAAB7nicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Bj04jGCeUCyhNnJbDJkdmaZ6RXCko/w4kERr36PN//GSbIHTSxoKKq66e6KUiks+v63t7a+sbm1Xdop7+7tHxxWjo5bVmeG8SbTUptORC2XQvEmCpS8kxpOk0jydjS+m/ntJ26s0OoRJykPEzpUIhaMopPaPY0i4bZfqfo1fw6ySoKCVKFAo1/56g00yxKukElqbTfwUwxzalAwyaflXmZ5StmYD | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 13 | 1/56g00yxKukElqbTfwUwxzalAwyaflXmZ5StmYDnnXUUXdkjCfnzsl504ZkFgbVwrJXP09kdPE2kkSuc6E4sguezPxP6+bYXwT5kKlGXLFFoviTBLUZPY7GQjDGcqJI5QZ4W4lbEQNZegSKrsQguWXV0nrshb4teDhqlq/LeIowSmcwQUEcA11uIcGNIHBGJ7hFd681Hvx3r2PReuaV8ycwB94nz+F64+t</latexit><latexit | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 14 | sha1_base64="f9ZxFFopka18yU5lyTwVt2fcbbg=">AAAB7nicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Bj04jGCeUCyhNnJbDJkdmaZ6RXCko/w4kERr36PN//GSbIHTSxoKKq66e6KUiks+v63t7a+sbm1Xdop7+7tHxxWjo5bVmeG8SbTUptORC2XQvEmCpS8kxpOk0jydjS+m/ntJ26s0OoRJykPEzpUIhaMopPaPY0i4bZfqfo1fw6ySoKCVKFAo1/56g00yxKukElqbTfwUwxzalAwyaflXmZ5StmYD | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 15 | 1/56g00yxKukElqbTfwUwxzalAwyaflXmZ5StmYDnnXUUXdkjCfnzsl504ZkFgbVwrJXP09kdPE2kkSuc6E4sguezPxP6+bYXwT5kKlGXLFFoviTBLUZPY7GQjDGcqJI5QZ4W4lbEQNZegSKrsQguWXV0nrshb4teDhqlq/LeIowSmcwQUEcA11uIcGNIHBGJ7hFd681Hvx3r2PReuaV8ycwB94nz+F64+t</latexit><latexit | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 16 | sha1_base64="f9ZxFFopka18yU5lyTwVt2fcbbg=">AAAB7nicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Bj04jGCeUCyhNnJbDJkdmaZ6RXCko/w4kERr36PN//GSbIHTSxoKKq66e6KUiks+v63t7a+sbm1Xdop7+7tHxxWjo5bVmeG8SbTUptORC2XQvEmCpS8kxpOk0jydjS+m/ntJ26s0OoRJykPEzpUIhaMopPaPY0i4bZfqfo1fw6ySoKCVKFAo1/56g00yxKukElqbTfwUwxzalAwyaflXmZ5StmYD | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 17 | 1/56g00yxKukElqbTfwUwxzalAwyaflXmZ5StmYDnnXUUXdkjCfnzsl504ZkFgbVwrJXP09kdPE2kkSuc6E4sguezPxP6+bYXwT5kKlGXLFFoviTBLUZPY7GQjDGcqJI5QZ4W4lbEQNZegSKrsQguWXV0nrshb4teDhqlq/LeIowSmcwQUEcA11uIcGNIHBGJ7hFd681Hvx3r2PReuaV8ycwB94nz+F64+t</latexit>time | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 18 | t+5time t-5time t-10 | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 19 | time ttime t+10time t-5time t+5time t+10time ttime t-10propagated instance segmentationfeature tensors for each frame in a video clip
time t-5time t+5time t+10time ttime t-10
Figure 2: An illustration of our MaskProp system, which takes as input a video clip centered around frame t, and outputs
a clip-level instance track. Our mask propagation framework can be summarized in three high-level steps: 1) An instancespecific feature map at time tis computed by masking the frame features at time twith the given instance segmentation
for framet(one for each instance detected in frame t). 2) Next, we use our mask propagation mechanism to temporally
propagate instance-specific features from frame tto all the other frames in the clip. 3) Lastly, our model predicts instancespecific segmentations in every frame of the clip by implicitly matching the propagated instance features with the frame-level
features computed at each time step. This last step yields clip-level instance tracks centered around frame t.
3) produces higher accuracy on YouTube-VIS [42].
3. Video Instance Segmentation
Problem Definition. Let us denote with V2RL3HW | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 20 | 3. Video Instance Segmentation
Problem Definition. Let us denote with V2RL3HW
an input video consisting of LRGB frames of spatial size
HW. The aim of our system is to segment and temporally
link all object instances that are visible for at least one frame
inVand that belong to a predefined set of categories C=
f1;:::;Kg. To achieve this goal, our model outputs a videolevel instance mask track Mi2RLHWwith a category
labelci2f1;:::;Kgand a confidence score si2[0;1]for
each object instance idetected in the video.
Evaluation Metric. Video instance segmentation is evaluated according to the metrics of average precision (AP) and
average recall (AR). Unlike in the image domain, these metrics are evaluated over the video sequence. Thus, to evaluate spatiotemporal consistency of the predicted mask sequences, the video Intersection over Union (IoU) between
a predicted object instance iand a ground truth object instancejis computed as:
IoU(i;j) =PT
t=1jMi(t)\~Mj(t)j
PT | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 21 | IoU(i;j) =PT
t=1jMi(t)\~Mj(t)j
PT
t=1jMi(t)[~Mj(t)j(1)
where ~Mj(t)is the ground-truth segmentation of object j
in framet. To achieve a large IoU, a model must not only
accurately classify and segment object instances at a framelevel, but also reliably track them over the video sequence.
As in the COCO benchmark for image segmentation [25], the metrics of AP and AR are computed separatelyfor each object category, and then averaged over ten IoU
thresholds from 50% to95% at increments of 5%. Lastly,
the resulting AP and AR metrics are averaged over the category set, which yields the final evaluation metric.
4. Mask Propagation
MaskProp takes a video Vof arbitrary length Las input and outputs video-level instance segmentation tracks
Mi, category labels ciand confidence scores sifor all
objectsidetected in the video. In order to achieve this
goal, our method first builds clip-level object instance tracks
Mi
t T:t+T2 R(2T+1)1HWfor each individual clip | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 22 | Mi
t T:t+T2 R(2T+1)1HWfor each individual clip
Vt T:t+T2R(2T+1)3HWof length (2T+ 1) in the
video, i.e., for t= 1;2;:::;L (clips at the beginning and
the end of the video will include fewer frames).
We want to use clips that are long enough to allow us
to jointly solve instance segmentation and tracking while
handling challenging cases of occlusion and motion blur.
At the same time, the clip should be short enough to allow
us to fit it at high spatial resolution in the memory of a GPU.
The resulting clip-level instance masks Mi
t T:t+Tproduced densely for all overlapping clips t= 1;:::;L are
then aggregated to produce video-level instance masks Mi.
Our approach for clip-level instance segmentation is described in subsections 4.1 and 4.2. We also illustrate it in
Figure 2. The subsequent clip-level instance mask aggregation method is presented in subsection 4.3.
Frame t
Bbox...
ClsMask
Frame t+... Mask Propagation CNN | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 23 | CNN <latexit | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 24 | sha1_base64="AU/vq+DAgMm1x9RhUc20aA7BJwY=">AAAB7XicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cK9gPaUDabTbt2kw27E6GE/gcvHhTx6v/x5r9x2+agrQ8GHu/NMDMvSKUw6LrfTmltfWNzq7xd2dnd2z+oHh61jco04y2mpNLdgBouRcJbKFDybqo5jQPJO8H4duZ3nrg2QiUPOEm5H9NhIiLBKFqp3Q+5RDqo1ty6OwdZJV5BalCgOah+9UPFspgnyCQ1pue5Kfo51SiY5NNKPzM8pWxMh | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 25 | ah+9UPFspgnyCQ1pue5Kfo51SiY5NNKPzM8pWxMh7xnaUJjbvx8fu2UnFklJJHSthIkc/X3RE5jYyZxYDtjiiOz7M3E/7xehtG1n4skzZAnbLEoyiRBRWavk1BozlBOLKFMC3srYSOqKUMbUMWG4C2/vEraF3XPrXv3l7XGTRFHGU7gFM7BgytowB00oQUMHuEZXuHNUc6L8+58LFpLTjFzDH/gfP4AkYyPHA==</latexit><latexit | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 26 | sha1_base64="AU/vq+DAgMm1x9RhUc20aA7BJwY=">AAAB7XicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cK9gPaUDabTbt2kw27E6GE/gcvHhTx6v/x5r9x2+agrQ8GHu/NMDMvSKUw6LrfTmltfWNzq7xd2dnd2z+oHh61jco04y2mpNLdgBouRcJbKFDybqo5jQPJO8H4duZ3nrg2QiUPOEm5H9NhIiLBKFqp3Q+5RDqo1ty6OwdZJV5BalCgOah+9UPFspgnyCQ1pue5Kfo51SiY5NNKPzM8pWxMh | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 27 | ah+9UPFspgnyCQ1pue5Kfo51SiY5NNKPzM8pWxMh7xnaUJjbvx8fu2UnFklJJHSthIkc/X3RE5jYyZxYDtjiiOz7M3E/7xehtG1n4skzZAnbLEoyiRBRWavk1BozlBOLKFMC3srYSOqKUMbUMWG4C2/vEraF3XPrXv3l7XGTRFHGU7gFM7BgytowB00oQUMHuEZXuHNUc6L8+58LFpLTjFzDH/gfP4AkYyPHA==</latexit><latexit | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 28 | sha1_base64="AU/vq+DAgMm1x9RhUc20aA7BJwY=">AAAB7XicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cK9gPaUDabTbt2kw27E6GE/gcvHhTx6v/x5r9x2+agrQ8GHu/NMDMvSKUw6LrfTmltfWNzq7xd2dnd2z+oHh61jco04y2mpNLdgBouRcJbKFDybqo5jQPJO8H4duZ3nrg2QiUPOEm5H9NhIiLBKFqp3Q+5RDqo1ty6OwdZJV5BalCgOah+9UPFspgnyCQ1pue5Kfo51SiY5NNKPzM8pWxMh | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 29 | ah+9UPFspgnyCQ1pue5Kfo51SiY5NNKPzM8pWxMh7xnaUJjbvx8fu2UnFklJJHSthIkc/X3RE5jYyZxYDtjiiOz7M3E/7xehtG1n4skzZAnbLEoyiRBRWavk1BozlBOLKFMC3srYSOqKUMbUMWG4C2/vEraF3XPrXv3l7XGTRFHGU7gFM7BgytowB00oQUMHuEZXuHNUc6L8+58LFpLTjFzDH/gfP4AkYyPHA==</latexit><latexit | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 30 | sha1_base64="AU/vq+DAgMm1x9RhUc20aA7BJwY=">AAAB7XicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cK9gPaUDabTbt2kw27E6GE/gcvHhTx6v/x5r9x2+agrQ8GHu/NMDMvSKUw6LrfTmltfWNzq7xd2dnd2z+oHh61jco04y2mpNLdgBouRcJbKFDybqo5jQPJO8H4duZ3nrg2QiUPOEm5H9NhIiLBKFqp3Q+5RDqo1ty6OwdZJV5BalCgOah+9UPFspgnyCQ1pue5Kfo51SiY5NNKPzM8pWxMh | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 31 | ah+9UPFspgnyCQ1pue5Kfo51SiY5NNKPzM8pWxMh7xnaUJjbvx8fu2UnFklJJHSthIkc/X3RE5jYyZxYDtjiiOz7M3E/7xehtG1n4skzZAnbLEoyiRBRWavk1BozlBOLKFMC3srYSOqKUMbUMWG4C2/vEraF3XPrXv3l7XGTRFHGU7gFM7BgytowB00oQUMHuEZXuHNUc6L8+58LFpLTjFzDH/gfP4AkYyPHA==</latexit> | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 32 | Figure 3: We adapt Mask R-CNN [16] to video by adding
a mask propagation branch, for tracking object segmentation instances in video. Given a video clip centered around
framet, our system outputs a clip-level instance segmentation track as well as a classification score and a bounding
box for each object instanced detected in frame t. For compactness, in this figure, we illustrate our system processing a
pair of frames but typically the propagation is applied from
the middle frame to all the other frames in the clip.
4.1. Video Mask R-CNN
Our video instance segmentation system is based on the
Mask R-CNN [16] model, which we adapt to video by
adding a mask propagation branch (See Figure 3). We train
our system with a multi-task loss Lt=Lcls
t+Lbox
t+
Lmask
t +Lprop
t wheretdenotes a time-step of a center
frame. We use identical loss terms Lcls
t;Lbox
t,Lmask
t as
in Mask R-CNN. The mask propagation loss is defined as:
Lprop
t =~NtX
it+TX
t0=t T1 sIoU (Mi
t T:t+T(t0);~Mi | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 33 | t =~NtX
it+TX
t0=t T1 sIoU (Mi
t T:t+T(t0);~Mi
t T:t+T(t0))(2)
whereMi
t T:t+T(t0)2[0;1]is the segmentation at time
t0for an instance ipredicted from a clip centered at tand
~Mi
t T:t+T(t0)is the corresponding ground truth mask at
timet0.~Ntis the number of ground truth object instances
in framet, andsIoU is defined as:
sIoU (A;B) =P
pA(p)B(p)P
pA(p) +B(p) A(p)B(p)(3)
where the summations in numerator and denominator are
performed over every pixel location p. The loss above is a
soft IoU loss, which we observed to work slightly better
than the standard cross entropy loss for our task.
4.2. Mask Propagation Branch
Overview. Our main technical contribution is the design of
a mask propagation branch, that allows our method to track | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 34 | 4.2. Mask Propagation Branch
Overview. Our main technical contribution is the design of
a mask propagation branch, that allows our method to track
object instances. Given a video clip Vt T:t+Tcenteredat framet, our system outputs clip-level instance masks
Mi
t T:t+Tfor each predicted object instance iin framet.
Our mask propagation branch can be described in three
high-level steps: 1) instance-specific feature computation,
2) temporal propagation of instance features, and 3) propagated instance segmentation. We will now describe each of
these steps in more detail. We introduce our mask propagation with the example of propagating object instance masks
from frametto framet+where2[ T:T].
Computing Instance Specific Features. The mask
branch of our model predicts frame-level instance masks
Mi
t2R1H0W0from single frame inputs. We then
use these frame-level instance masks to compute instancespecific features for frame t. Specifically, for each object instancei, we compute an element-wise product between Mi
t
and the feature tensor from the backbone network ft. This
then yields a set of new feature tensors fi | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 35 | t
and the feature tensor from the backbone network ft. This
then yields a set of new feature tensors fi
t2RCH0W0,
wherei= 1;::;Nt, andNtis the number of object instances
detected in frame t. In other words, for each object instance
i, we are zeroing out the feature values in ftthat correspond
to pixels not belonging to that object instance.
Temporally Propagating Instance Features. Given
frame-level features ft;ft+2RCH0W0and instancespecific feature tensor fi
t, our method generates a propagated instance feature tensor gi
t;t+. Intuitively, gi
t;t+represents the features predicted by our model for object instanceiin framet+from an instance-specific feature tensorfi
t. The tensor gi
t;t+is generated by warping features
fi
tusing the alignment computed from frame-level features
ftandft+. We implement the propagation mechanism via
a deformable convolution [12], which has previously been
used for aligning features computed from separate frames | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 36 | ftandft+. We implement the propagation mechanism via
a deformable convolution [12], which has previously been
used for aligning features computed from separate frames
of a video [5, 3]. Specifically, we compute the elementwise difference of tensors ft;ft+and feed it through a
simple residual block [17], which predicts motion offsets
ot;t+2R2k2H0W0. These offsets contain (x;y)sampling locations for each entry of a kkdeformable convolution kernel [12]. The propagation step takes as inputs 1)
the offsetsot;t+and 2) the instance feature tensor fi
t, and
then applies deformable convolution to output the propagated instance feature tensor gi
t;t+for each instance i. We
use subscript t;t+to denote the propagated instance feature because, although gis obtained by propagating the feature tensorfi
t, the offset computation uses both frame tand
framet+. We stress that no explicit ground truth alignment is available between frames. The deformable convolutional kernels are supervised implicitly by optimizing Eq. 2.
Segmenting Propagated Instances. Lastly, we use our
propagated feature map gi | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 37 | Segmenting Propagated Instances. Lastly, we use our
propagated feature map gi
t;t+for predicting a corresponding object instance mask in frame t+. To do this we
first, construct a new feature tensor i
t;t+=gi
t;t++ft+.
The addition effectively overimposes the tensor gi
t;t+predicted from time tfor object instance iin framet+, with
Tensor t | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 38 | <latexit | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 39 | sha1_base64="xbJVLRy5zV8jtLsjAis5DXwMT4U=">AAAB7nicbVDJSgNBEK1xjXGLevTSGARPYUYEPQa9eIxgFkiG0NPpSZr0MvQihCEf4cWDIl79Hm/+jZ1kDpr4oODxXhVV9ZKMM2PD8DtYW9/Y3Nou7ZR39/YPDitHxy2jnCa0SRRXupNgQzmTtGmZ5bSTaYpFwmk7Gd/N/PYT1YYp+WgnGY0FHkqWMoKtl9o9JZh0pl+phrVwDrRKooJUoUCjX/nqDRRxgkpLODamG4WZjXOsLSOcTss9Z2iGyRgPa | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 40 | /nqDRRxgkpLODamG4WZjXOsLSOcTss9Z2iGyRgPaddTiQU1cT4/d4rOvTJAqdK+pEVz9fdEjoUxE5H4ToHtyCx7M/E/r+tsehPnTGbOUkkWi1LHkVVo9jsaME2J5RNPMNHM34rICGtMrE+o7EOIll9eJa3LWhTWooerav22iKMEp3AGFxDBNdThHhrQBAJjeIZXeAuy4CV4Dz4WrWtBMXMCfxB8/gCVCY+3</latexit><latexit | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 41 | sha1_base64="xbJVLRy5zV8jtLsjAis5DXwMT4U=">AAAB7nicbVDJSgNBEK1xjXGLevTSGARPYUYEPQa9eIxgFkiG0NPpSZr0MvQihCEf4cWDIl79Hm/+jZ1kDpr4oODxXhVV9ZKMM2PD8DtYW9/Y3Nou7ZR39/YPDitHxy2jnCa0SRRXupNgQzmTtGmZ5bSTaYpFwmk7Gd/N/PYT1YYp+WgnGY0FHkqWMoKtl9o9JZh0pl+phrVwDrRKooJUoUCjX/nqDRRxgkpLODamG4WZjXOsLSOcTss9Z2iGyRgPa | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 42 | /nqDRRxgkpLODamG4WZjXOsLSOcTss9Z2iGyRgPaddTiQU1cT4/d4rOvTJAqdK+pEVz9fdEjoUxE5H4ToHtyCx7M/E/r+tsehPnTGbOUkkWi1LHkVVo9jsaME2J5RNPMNHM34rICGtMrE+o7EOIll9eJa3LWhTWooerav22iKMEp3AGFxDBNdThHhrQBAJjeIZXeAuy4CV4Dz4WrWtBMXMCfxB8/gCVCY+3</latexit><latexit | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 43 | sha1_base64="xbJVLRy5zV8jtLsjAis5DXwMT4U=">AAAB7nicbVDJSgNBEK1xjXGLevTSGARPYUYEPQa9eIxgFkiG0NPpSZr0MvQihCEf4cWDIl79Hm/+jZ1kDpr4oODxXhVV9ZKMM2PD8DtYW9/Y3Nou7ZR39/YPDitHxy2jnCa0SRRXupNgQzmTtGmZ5bSTaYpFwmk7Gd/N/PYT1YYp+WgnGY0FHkqWMoKtl9o9JZh0pl+phrVwDrRKooJUoUCjX/nqDRRxgkpLODamG4WZjXOsLSOcTss9Z2iGyRgPa | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 44 | /nqDRRxgkpLODamG4WZjXOsLSOcTss9Z2iGyRgPaddTiQU1cT4/d4rOvTJAqdK+pEVz9fdEjoUxE5H4ToHtyCx7M/E/r+tsehPnTGbOUkkWi1LHkVVo9jsaME2J5RNPMNHM34rICGtMrE+o7EOIll9eJa3LWhTWooerav22iKMEp3AGFxDBNdThHhrQBAJjeIZXeAuy4CV4Dz4WrWtBMXMCfxB8/gCVCY+3</latexit><latexit | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 45 | sha1_base64="xbJVLRy5zV8jtLsjAis5DXwMT4U=">AAAB7nicbVDJSgNBEK1xjXGLevTSGARPYUYEPQa9eIxgFkiG0NPpSZr0MvQihCEf4cWDIl79Hm/+jZ1kDpr4oODxXhVV9ZKMM2PD8DtYW9/Y3Nou7ZR39/YPDitHxy2jnCa0SRRXupNgQzmTtGmZ5bSTaYpFwmk7Gd/N/PYT1YYp+WgnGY0FHkqWMoKtl9o9JZh0pl+phrVwDrRKooJUoUCjX/nqDRRxgkpLODamG4WZjXOsLSOcTss9Z2iGyRgPa | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 46 | /nqDRRxgkpLODamG4WZjXOsLSOcTss9Z2iGyRgPaddTiQU1cT4/d4rOvTJAqdK+pEVz9fdEjoUxE5H4ToHtyCx7M/E/r+tsehPnTGbOUkkWi1LHkVVo9jsaME2J5RNPMNHM34rICGtMrE+o7EOIll9eJa3LWhTWooerav22iKMEp3AGFxDBNdThHhrQBAJjeIZXeAuy4CV4Dz4WrWtBMXMCfxB8/gCVCY+3</latexit>Tensor | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 47 | t+ <latexit | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 48 | sha1_base64="AU/vq+DAgMm1x9RhUc20aA7BJwY=">AAAB7XicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cK9gPaUDabTbt2kw27E6GE/gcvHhTx6v/x5r9x2+agrQ8GHu/NMDMvSKUw6LrfTmltfWNzq7xd2dnd2z+oHh61jco04y2mpNLdgBouRcJbKFDybqo5jQPJO8H4duZ3nrg2QiUPOEm5H9NhIiLBKFqp3Q+5RDqo1ty6OwdZJV5BalCgOah+9UPFspgnyCQ1pue5Kfo51SiY5NNKPzM8pWxMh | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 49 | ah+9UPFspgnyCQ1pue5Kfo51SiY5NNKPzM8pWxMh7xnaUJjbvx8fu2UnFklJJHSthIkc/X3RE5jYyZxYDtjiiOz7M3E/7xehtG1n4skzZAnbLEoyiRBRWavk1BozlBOLKFMC3srYSOqKUMbUMWG4C2/vEraF3XPrXv3l7XGTRFHGU7gFM7BgytowB00oQUMHuEZXuHNUc6L8+58LFpLTjFzDH/gfP4AkYyPHA==</latexit><latexit | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 50 | sha1_base64="AU/vq+DAgMm1x9RhUc20aA7BJwY=">AAAB7XicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cK9gPaUDabTbt2kw27E6GE/gcvHhTx6v/x5r9x2+agrQ8GHu/NMDMvSKUw6LrfTmltfWNzq7xd2dnd2z+oHh61jco04y2mpNLdgBouRcJbKFDybqo5jQPJO8H4duZ3nrg2QiUPOEm5H9NhIiLBKFqp3Q+5RDqo1ty6OwdZJV5BalCgOah+9UPFspgnyCQ1pue5Kfo51SiY5NNKPzM8pWxMh | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 51 | ah+9UPFspgnyCQ1pue5Kfo51SiY5NNKPzM8pWxMh7xnaUJjbvx8fu2UnFklJJHSthIkc/X3RE5jYyZxYDtjiiOz7M3E/7xehtG1n4skzZAnbLEoyiRBRWavk1BozlBOLKFMC3srYSOqKUMbUMWG4C2/vEraF3XPrXv3l7XGTRFHGU7gFM7BgytowB00oQUMHuEZXuHNUc6L8+58LFpLTjFzDH/gfP4AkYyPHA==</latexit><latexit | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 52 | sha1_base64="AU/vq+DAgMm1x9RhUc20aA7BJwY=">AAAB7XicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cK9gPaUDabTbt2kw27E6GE/gcvHhTx6v/x5r9x2+agrQ8GHu/NMDMvSKUw6LrfTmltfWNzq7xd2dnd2z+oHh61jco04y2mpNLdgBouRcJbKFDybqo5jQPJO8H4duZ3nrg2QiUPOEm5H9NhIiLBKFqp3Q+5RDqo1ty6OwdZJV5BalCgOah+9UPFspgnyCQ1pue5Kfo51SiY5NNKPzM8pWxMh | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 53 | ah+9UPFspgnyCQ1pue5Kfo51SiY5NNKPzM8pWxMh7xnaUJjbvx8fu2UnFklJJHSthIkc/X3RE5jYyZxYDtjiiOz7M3E/7xehtG1n4skzZAnbLEoyiRBRWavk1BozlBOLKFMC3srYSOqKUMbUMWG4C2/vEraF3XPrXv3l7XGTRFHGU7gFM7BgytowB00oQUMHuEZXuHNUc6L8+58LFpLTjFzDH/gfP4AkYyPHA==</latexit><latexit | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 54 | sha1_base64="AU/vq+DAgMm1x9RhUc20aA7BJwY=">AAAB7XicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cK9gPaUDabTbt2kw27E6GE/gcvHhTx6v/x5r9x2+agrQ8GHu/NMDMvSKUw6LrfTmltfWNzq7xd2dnd2z+oHh61jco04y2mpNLdgBouRcJbKFDybqo5jQPJO8H4duZ3nrg2QiUPOEm5H9NhIiLBKFqp3Q+5RDqo1ty6OwdZJV5BalCgOah+9UPFspgnyCQ1pue5Kfo51SiY5NNKPzM8pWxMh | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 55 | ah+9UPFspgnyCQ1pue5Kfo51SiY5NNKPzM8pWxMh7xnaUJjbvx8fu2UnFklJJHSthIkc/X3RE5jYyZxYDtjiiOz7M3E/7xehtG1n4skzZAnbLEoyiRBRWavk1BozlBOLKFMC3srYSOqKUMbUMWG4C2/vEraF3XPrXv3l7XGTRFHGU7gFM7BgytowB00oQUMHuEZXuHNUc6L8+58LFpLTjFzDH/gfP4AkYyPHA==</latexit> | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 56 | Residual Block | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 57 | ⌦<latexit | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 58 | sha1_base64="f9ZxFFopka18yU5lyTwVt2fcbbg=">AAAB7nicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Bj04jGCeUCyhNnJbDJkdmaZ6RXCko/w4kERr36PN//GSbIHTSxoKKq66e6KUiks+v63t7a+sbm1Xdop7+7tHxxWjo5bVmeG8SbTUptORC2XQvEmCpS8kxpOk0jydjS+m/ntJ26s0OoRJykPEzpUIhaMopPaPY0i4bZfqfo1fw6ySoKCVKFAo1/56g00yxKukElqbTfwUwxzalAwyaflXmZ5StmYD | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 59 | 1/56g00yxKukElqbTfwUwxzalAwyaflXmZ5StmYDnnXUUXdkjCfnzsl504ZkFgbVwrJXP09kdPE2kkSuc6E4sguezPxP6+bYXwT5kKlGXLFFoviTBLUZPY7GQjDGcqJI5QZ4W4lbEQNZegSKrsQguWXV0nrshb4teDhqlq/LeIowSmcwQUEcA11uIcGNIHBGJ7hFd681Hvx3r2PReuaV8ycwB94nz+F64+t</latexit><latexit | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 60 | sha1_base64="f9ZxFFopka18yU5lyTwVt2fcbbg=">AAAB7nicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Bj04jGCeUCyhNnJbDJkdmaZ6RXCko/w4kERr36PN//GSbIHTSxoKKq66e6KUiks+v63t7a+sbm1Xdop7+7tHxxWjo5bVmeG8SbTUptORC2XQvEmCpS8kxpOk0jydjS+m/ntJ26s0OoRJykPEzpUIhaMopPaPY0i4bZfqfo1fw6ySoKCVKFAo1/56g00yxKukElqbTfwUwxzalAwyaflXmZ5StmYD | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 61 | 1/56g00yxKukElqbTfwUwxzalAwyaflXmZ5StmYDnnXUUXdkjCfnzsl504ZkFgbVwrJXP09kdPE2kkSuc6E4sguezPxP6+bYXwT5kKlGXLFFoviTBLUZPY7GQjDGcqJI5QZ4W4lbEQNZegSKrsQguWXV0nrshb4teDhqlq/LeIowSmcwQUEcA11uIcGNIHBGJ7hFd681Hvx3r2PReuaV8ycwB94nz+F64+t</latexit><latexit | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 62 | sha1_base64="f9ZxFFopka18yU5lyTwVt2fcbbg=">AAAB7nicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Bj04jGCeUCyhNnJbDJkdmaZ6RXCko/w4kERr36PN//GSbIHTSxoKKq66e6KUiks+v63t7a+sbm1Xdop7+7tHxxWjo5bVmeG8SbTUptORC2XQvEmCpS8kxpOk0jydjS+m/ntJ26s0OoRJykPEzpUIhaMopPaPY0i4bZfqfo1fw6ySoKCVKFAo1/56g00yxKukElqbTfwUwxzalAwyaflXmZ5StmYD | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 63 | 1/56g00yxKukElqbTfwUwxzalAwyaflXmZ5StmYDnnXUUXdkjCfnzsl504ZkFgbVwrJXP09kdPE2kkSuc6E4sguezPxP6+bYXwT5kKlGXLFFoviTBLUZPY7GQjDGcqJI5QZ4W4lbEQNZegSKrsQguWXV0nrshb4teDhqlq/LeIowSmcwQUEcA11uIcGNIHBGJ7hFd681Hvx3r2PReuaV8ycwB94nz+F64+t</latexit><latexit | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 64 | sha1_base64="f9ZxFFopka18yU5lyTwVt2fcbbg=">AAAB7nicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Bj04jGCeUCyhNnJbDJkdmaZ6RXCko/w4kERr36PN//GSbIHTSxoKKq66e6KUiks+v63t7a+sbm1Xdop7+7tHxxWjo5bVmeG8SbTUptORC2XQvEmCpS8kxpOk0jydjS+m/ntJ26s0OoRJykPEzpUIhaMopPaPY0i4bZfqfo1fw6ySoKCVKFAo1/56g00yxKukElqbTfwUwxzalAwyaflXmZ5StmYD | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 65 | 1/56g00yxKukElqbTfwUwxzalAwyaflXmZ5StmYDnnXUUXdkjCfnzsl504ZkFgbVwrJXP09kdPE2kkSuc6E4sguezPxP6+bYXwT5kKlGXLFFoviTBLUZPY7GQjDGcqJI5QZ4W4lbEQNZegSKrsQguWXV0nrshb4teDhqlq/LeIowSmcwQUEcA11uIcGNIHBGJ7hFd681Hvx3r2PReuaV8ycwB94nz+F64+t</latexit>⌦<latexit | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 66 | sha1_base64="f9ZxFFopka18yU5lyTwVt2fcbbg=">AAAB7nicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Bj04jGCeUCyhNnJbDJkdmaZ6RXCko/w4kERr36PN//GSbIHTSxoKKq66e6KUiks+v63t7a+sbm1Xdop7+7tHxxWjo5bVmeG8SbTUptORC2XQvEmCpS8kxpOk0jydjS+m/ntJ26s0OoRJykPEzpUIhaMopPaPY0i4bZfqfo1fw6ySoKCVKFAo1/56g00yxKukElqbTfwUwxzalAwyaflXmZ5StmYD | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 67 | 1/56g00yxKukElqbTfwUwxzalAwyaflXmZ5StmYDnnXUUXdkjCfnzsl504ZkFgbVwrJXP09kdPE2kkSuc6E4sguezPxP6+bYXwT5kKlGXLFFoviTBLUZPY7GQjDGcqJI5QZ4W4lbEQNZegSKrsQguWXV0nrshb4teDhqlq/LeIowSmcwQUEcA11uIcGNIHBGJ7hFd681Hvx3r2PReuaV8ycwB94nz+F64+t</latexit><latexit | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 68 | sha1_base64="f9ZxFFopka18yU5lyTwVt2fcbbg=">AAAB7nicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Bj04jGCeUCyhNnJbDJkdmaZ6RXCko/w4kERr36PN//GSbIHTSxoKKq66e6KUiks+v63t7a+sbm1Xdop7+7tHxxWjo5bVmeG8SbTUptORC2XQvEmCpS8kxpOk0jydjS+m/ntJ26s0OoRJykPEzpUIhaMopPaPY0i4bZfqfo1fw6ySoKCVKFAo1/56g00yxKukElqbTfwUwxzalAwyaflXmZ5StmYD | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 69 | 1/56g00yxKukElqbTfwUwxzalAwyaflXmZ5StmYDnnXUUXdkjCfnzsl504ZkFgbVwrJXP09kdPE2kkSuc6E4sguezPxP6+bYXwT5kKlGXLFFoviTBLUZPY7GQjDGcqJI5QZ4W4lbEQNZegSKrsQguWXV0nrshb4teDhqlq/LeIowSmcwQUEcA11uIcGNIHBGJ7hFd681Hvx3r2PReuaV8ycwB94nz+F64+t</latexit><latexit | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 70 | sha1_base64="f9ZxFFopka18yU5lyTwVt2fcbbg=">AAAB7nicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Bj04jGCeUCyhNnJbDJkdmaZ6RXCko/w4kERr36PN//GSbIHTSxoKKq66e6KUiks+v63t7a+sbm1Xdop7+7tHxxWjo5bVmeG8SbTUptORC2XQvEmCpS8kxpOk0jydjS+m/ntJ26s0OoRJykPEzpUIhaMopPaPY0i4bZfqfo1fw6ySoKCVKFAo1/56g00yxKukElqbTfwUwxzalAwyaflXmZ5StmYD | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 71 | 1/56g00yxKukElqbTfwUwxzalAwyaflXmZ5StmYDnnXUUXdkjCfnzsl504ZkFgbVwrJXP09kdPE2kkSuc6E4sguezPxP6+bYXwT5kKlGXLFFoviTBLUZPY7GQjDGcqJI5QZ4W4lbEQNZegSKrsQguWXV0nrshb4teDhqlq/LeIowSmcwQUEcA11uIcGNIHBGJ7hFd681Hvx3r2PReuaV8ycwB94nz+F64+t</latexit><latexit | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 72 | sha1_base64="f9ZxFFopka18yU5lyTwVt2fcbbg=">AAAB7nicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Bj04jGCeUCyhNnJbDJkdmaZ6RXCko/w4kERr36PN//GSbIHTSxoKKq66e6KUiks+v63t7a+sbm1Xdop7+7tHxxWjo5bVmeG8SbTUptORC2XQvEmCpS8kxpOk0jydjS+m/ntJ26s0OoRJykPEzpUIhaMopPaPY0i4bZfqfo1fw6ySoKCVKFAo1/56g00yxKukElqbTfwUwxzalAwyaflXmZ5StmYD | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 73 | 1/56g00yxKukElqbTfwUwxzalAwyaflXmZ5StmYDnnXUUXdkjCfnzsl504ZkFgbVwrJXP09kdPE2kkSuc6E4sguezPxP6+bYXwT5kKlGXLFFoviTBLUZPY7GQjDGcqJI5QZ4W4lbEQNZegSKrsQguWXV0nrshb4teDhqlq/LeIowSmcwQUEcA11uIcGNIHBGJ7hFd681Hvx3r2PReuaV8ycwB94nz+F64+t</latexit> | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 74 | Deform. Conv. Offsets <latexit | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 75 | sha1_base64="N4ztu7rA6GcWPA6nADZ8tMOcvqE=">AAAB7XicbVDLSgNBEOyNrxhfUY9eBoPgKeyKoMegF48RzAOSJcxOZpMxszPLTK8QQv7BiwdFvPo/3vwbJ8keNLGgoajqprsrSqWw6PvfXmFtfWNzq7hd2tnd2z8oHx41rc4M4w2mpTbtiFouheINFCh5OzWcJpHkrWh0O/NbT9xYodUDjlMeJnSgRCwYRSc1uzqVme2VK37Vn4OskiAnFchR75W/un3NsoQrZJJa2wn8FMMJNSiY5NNSN7M8pWxEB | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 76 | 5W/un3NsoQrZJJa2wn8FMMJNSiY5NNSN7M8pWxEB7zjqKIJt+Fkfu2UnDmlT2JtXCkkc/X3xIQm1o6TyHUmFId22ZuJ/3mdDOPrcCJUmiFXbLEoziRBTWavk74wnKEcO0KZEe5WwobUUIYuoJILIVh+eZU0L6qBXw3uLyu1mzyOIpzAKZxDAFdQgzuoQwMYPMIzvMKbp70X7937WLQWvHzmGP7A+/wBz/6PRQ==</latexit><latexit | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 77 | sha1_base64="N4ztu7rA6GcWPA6nADZ8tMOcvqE=">AAAB7XicbVDLSgNBEOyNrxhfUY9eBoPgKeyKoMegF48RzAOSJcxOZpMxszPLTK8QQv7BiwdFvPo/3vwbJ8keNLGgoajqprsrSqWw6PvfXmFtfWNzq7hd2tnd2z8oHx41rc4M4w2mpTbtiFouheINFCh5OzWcJpHkrWh0O/NbT9xYodUDjlMeJnSgRCwYRSc1uzqVme2VK37Vn4OskiAnFchR75W/un3NsoQrZJJa2wn8FMMJNSiY5NNSN7M8pWxEB | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 78 | 5W/un3NsoQrZJJa2wn8FMMJNSiY5NNSN7M8pWxEB7zjqKIJt+Fkfu2UnDmlT2JtXCkkc/X3xIQm1o6TyHUmFId22ZuJ/3mdDOPrcCJUmiFXbLEoziRBTWavk74wnKEcO0KZEe5WwobUUIYuoJILIVh+eZU0L6qBXw3uLyu1mzyOIpzAKZxDAFdQgzuoQwMYPMIzvMKbp70X7937WLQWvHzmGP7A+/wBz/6PRQ==</latexit><latexit | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 79 | sha1_base64="N4ztu7rA6GcWPA6nADZ8tMOcvqE=">AAAB7XicbVDLSgNBEOyNrxhfUY9eBoPgKeyKoMegF48RzAOSJcxOZpMxszPLTK8QQv7BiwdFvPo/3vwbJ8keNLGgoajqprsrSqWw6PvfXmFtfWNzq7hd2tnd2z8oHx41rc4M4w2mpTbtiFouheINFCh5OzWcJpHkrWh0O/NbT9xYodUDjlMeJnSgRCwYRSc1uzqVme2VK37Vn4OskiAnFchR75W/un3NsoQrZJJa2wn8FMMJNSiY5NNSN7M8pWxEB | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 80 | 5W/un3NsoQrZJJa2wn8FMMJNSiY5NNSN7M8pWxEB7zjqKIJt+Fkfu2UnDmlT2JtXCkkc/X3xIQm1o6TyHUmFId22ZuJ/3mdDOPrcCJUmiFXbLEoziRBTWavk74wnKEcO0KZEe5WwobUUIYuoJILIVh+eZU0L6qBXw3uLyu1mzyOIpzAKZxDAFdQgzuoQwMYPMIzvMKbp70X7937WLQWvHzmGP7A+/wBz/6PRQ==</latexit><latexit | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 81 | sha1_base64="N4ztu7rA6GcWPA6nADZ8tMOcvqE=">AAAB7XicbVDLSgNBEOyNrxhfUY9eBoPgKeyKoMegF48RzAOSJcxOZpMxszPLTK8QQv7BiwdFvPo/3vwbJ8keNLGgoajqprsrSqWw6PvfXmFtfWNzq7hd2tnd2z8oHx41rc4M4w2mpTbtiFouheINFCh5OzWcJpHkrWh0O/NbT9xYodUDjlMeJnSgRCwYRSc1uzqVme2VK37Vn4OskiAnFchR75W/un3NsoQrZJJa2wn8FMMJNSiY5NNSN7M8pWxEB | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 82 | 5W/un3NsoQrZJJa2wn8FMMJNSiY5NNSN7M8pWxEB7zjqKIJt+Fkfu2UnDmlT2JtXCkkc/X3xIQm1o6TyHUmFId22ZuJ/3mdDOPrcCJUmiFXbLEoziRBTWavk74wnKEcO0KZEe5WwobUUIYuoJILIVh+eZU0L6qBXw3uLyu1mzyOIpzAKZxDAFdQgzuoQwMYPMIzvMKbp70X7937WLQWvHzmGP7A+/wBz/6PRQ==</latexit> <latexit | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 83 | sha1_base64="N4ztu7rA6GcWPA6nADZ8tMOcvqE=">AAAB7XicbVDLSgNBEOyNrxhfUY9eBoPgKeyKoMegF48RzAOSJcxOZpMxszPLTK8QQv7BiwdFvPo/3vwbJ8keNLGgoajqprsrSqWw6PvfXmFtfWNzq7hd2tnd2z8oHx41rc4M4w2mpTbtiFouheINFCh5OzWcJpHkrWh0O/NbT9xYodUDjlMeJnSgRCwYRSc1uzqVme2VK37Vn4OskiAnFchR75W/un3NsoQrZJJa2wn8FMMJNSiY5NNSN7M8pWxEB | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 84 | 5W/un3NsoQrZJJa2wn8FMMJNSiY5NNSN7M8pWxEB7zjqKIJt+Fkfu2UnDmlT2JtXCkkc/X3xIQm1o6TyHUmFId22ZuJ/3mdDOPrcCJUmiFXbLEoziRBTWavk74wnKEcO0KZEe5WwobUUIYuoJILIVh+eZU0L6qBXw3uLyu1mzyOIpzAKZxDAFdQgzuoQwMYPMIzvMKbp70X7937WLQWvHzmGP7A+/wBz/6PRQ==</latexit><latexit | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 85 | sha1_base64="N4ztu7rA6GcWPA6nADZ8tMOcvqE=">AAAB7XicbVDLSgNBEOyNrxhfUY9eBoPgKeyKoMegF48RzAOSJcxOZpMxszPLTK8QQv7BiwdFvPo/3vwbJ8keNLGgoajqprsrSqWw6PvfXmFtfWNzq7hd2tnd2z8oHx41rc4M4w2mpTbtiFouheINFCh5OzWcJpHkrWh0O/NbT9xYodUDjlMeJnSgRCwYRSc1uzqVme2VK37Vn4OskiAnFchR75W/un3NsoQrZJJa2wn8FMMJNSiY5NNSN7M8pWxEB | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 86 | 5W/un3NsoQrZJJa2wn8FMMJNSiY5NNSN7M8pWxEB7zjqKIJt+Fkfu2UnDmlT2JtXCkkc/X3xIQm1o6TyHUmFId22ZuJ/3mdDOPrcCJUmiFXbLEoziRBTWavk74wnKEcO0KZEe5WwobUUIYuoJILIVh+eZU0L6qBXw3uLyu1mzyOIpzAKZxDAFdQgzuoQwMYPMIzvMKbp70X7937WLQWvHzmGP7A+/wBz/6PRQ==</latexit><latexit | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 87 | sha1_base64="N4ztu7rA6GcWPA6nADZ8tMOcvqE=">AAAB7XicbVDLSgNBEOyNrxhfUY9eBoPgKeyKoMegF48RzAOSJcxOZpMxszPLTK8QQv7BiwdFvPo/3vwbJ8keNLGgoajqprsrSqWw6PvfXmFtfWNzq7hd2tnd2z8oHx41rc4M4w2mpTbtiFouheINFCh5OzWcJpHkrWh0O/NbT9xYodUDjlMeJnSgRCwYRSc1uzqVme2VK37Vn4OskiAnFchR75W/un3NsoQrZJJa2wn8FMMJNSiY5NNSN7M8pWxEB | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 88 | 5W/un3NsoQrZJJa2wn8FMMJNSiY5NNSN7M8pWxEB7zjqKIJt+Fkfu2UnDmlT2JtXCkkc/X3xIQm1o6TyHUmFId22ZuJ/3mdDOPrcCJUmiFXbLEoziRBTWavk74wnKEcO0KZEe5WwobUUIYuoJILIVh+eZU0L6qBXw3uLyu1mzyOIpzAKZxDAFdQgzuoQwMYPMIzvMKbp70X7937WLQWvHzmGP7A+/wBz/6PRQ==</latexit><latexit | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 89 | sha1_base64="N4ztu7rA6GcWPA6nADZ8tMOcvqE=">AAAB7XicbVDLSgNBEOyNrxhfUY9eBoPgKeyKoMegF48RzAOSJcxOZpMxszPLTK8QQv7BiwdFvPo/3vwbJ8keNLGgoajqprsrSqWw6PvfXmFtfWNzq7hd2tnd2z8oHx41rc4M4w2mpTbtiFouheINFCh5OzWcJpHkrWh0O/NbT9xYodUDjlMeJnSgRCwYRSc1uzqVme2VK37Vn4OskiAnFchR75W/un3NsoQrZJJa2wn8FMMJNSiY5NNSN7M8pWxEB | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 90 | 5W/un3NsoQrZJJa2wn8FMMJNSiY5NNSN7M8pWxEB7zjqKIJt+Fkfu2UnDmlT2JtXCkkc/X3xIQm1o6TyHUmFId22ZuJ/3mdDOPrcCJUmiFXbLEoziRBTWavk74wnKEcO0KZEe5WwobUUIYuoJILIVh+eZU0L6qBXw3uLyu1mzyOIpzAKZxDAFdQgzuoQwMYPMIzvMKbp70X7937WLQWvHzmGP7A+/wBz/6PRQ==</latexit> | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 91 | 2D Conv. | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 92 | Softmax Attention⌦<latexit | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 93 | sha1_base64="f9ZxFFopka18yU5lyTwVt2fcbbg=">AAAB7nicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Bj04jGCeUCyhNnJbDJkdmaZ6RXCko/w4kERr36PN//GSbIHTSxoKKq66e6KUiks+v63t7a+sbm1Xdop7+7tHxxWjo5bVmeG8SbTUptORC2XQvEmCpS8kxpOk0jydjS+m/ntJ26s0OoRJykPEzpUIhaMopPaPY0i4bZfqfo1fw6ySoKCVKFAo1/56g00yxKukElqbTfwUwxzalAwyaflXmZ5StmYD | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 94 | 1/56g00yxKukElqbTfwUwxzalAwyaflXmZ5StmYDnnXUUXdkjCfnzsl504ZkFgbVwrJXP09kdPE2kkSuc6E4sguezPxP6+bYXwT5kKlGXLFFoviTBLUZPY7GQjDGcqJI5QZ4W4lbEQNZegSKrsQguWXV0nrshb4teDhqlq/LeIowSmcwQUEcA11uIcGNIHBGJ7hFd681Hvx3r2PReuaV8ycwB94nz+F64+t</latexit><latexit | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 95 | sha1_base64="f9ZxFFopka18yU5lyTwVt2fcbbg=">AAAB7nicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Bj04jGCeUCyhNnJbDJkdmaZ6RXCko/w4kERr36PN//GSbIHTSxoKKq66e6KUiks+v63t7a+sbm1Xdop7+7tHxxWjo5bVmeG8SbTUptORC2XQvEmCpS8kxpOk0jydjS+m/ntJ26s0OoRJykPEzpUIhaMopPaPY0i4bZfqfo1fw6ySoKCVKFAo1/56g00yxKukElqbTfwUwxzalAwyaflXmZ5StmYD | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 96 | 1/56g00yxKukElqbTfwUwxzalAwyaflXmZ5StmYDnnXUUXdkjCfnzsl504ZkFgbVwrJXP09kdPE2kkSuc6E4sguezPxP6+bYXwT5kKlGXLFFoviTBLUZPY7GQjDGcqJI5QZ4W4lbEQNZegSKrsQguWXV0nrshb4teDhqlq/LeIowSmcwQUEcA11uIcGNIHBGJ7hFd681Hvx3r2PReuaV8ycwB94nz+F64+t</latexit><latexit | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 97 | sha1_base64="f9ZxFFopka18yU5lyTwVt2fcbbg=">AAAB7nicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Bj04jGCeUCyhNnJbDJkdmaZ6RXCko/w4kERr36PN//GSbIHTSxoKKq66e6KUiks+v63t7a+sbm1Xdop7+7tHxxWjo5bVmeG8SbTUptORC2XQvEmCpS8kxpOk0jydjS+m/ntJ26s0OoRJykPEzpUIhaMopPaPY0i4bZfqfo1fw6ySoKCVKFAo1/56g00yxKukElqbTfwUwxzalAwyaflXmZ5StmYD | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
1912.04573 | 98 | 1/56g00yxKukElqbTfwUwxzalAwyaflXmZ5StmYDnnXUUXdkjCfnzsl504ZkFgbVwrJXP09kdPE2kkSuc6E4sguezPxP6+bYXwT5kKlGXLFFoviTBLUZPY7GQjDGcqJI5QZ4W4lbEQNZegSKrsQguWXV0nrshb4teDhqlq/LeIowSmcwQUEcA11uIcGNIHBGJ7hFd681Hvx3r2PReuaV8ycwB94nz+F64+t</latexit><latexit | 1912.04573 | Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation | We introduce a method for simultaneously classifying, segmenting and tracking
object instances in a video sequence. Our method, named MaskProp, adapts the
popular Mask R-CNN to video by adding a mask propagation branch that propagates
frame-level object instance masks from each video frame to all the other frames
in a video clip. This allows our system to predict clip-level instance tracks
with respect to the object instances segmented in the middle frame of the clip.
Clip-level instance tracks generated densely for each frame in the sequence are
finally aggregated to produce video-level object instance segmentation and
classification. Our experiments demonstrate that our clip-level instance
segmentation makes our approach robust to motion blur and object occlusions in
video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset,
outperforming the ICCV 2019 video instance segmentation challenge winner
despite being much simpler and using orders of magnitude less labeled data
(1.3M vs 1B images and 860K vs 14M bounding boxes). | http://arxiv.org/pdf/1912.04573 | [
"Gedas Bertasius",
"Lorenzo Torresani"
] | [
"cs.CV"
] | CVPR 2020 Best Paper Nominee | null | cs.CV | 20191210 | 20210709 | [
{
"id": "1704.02386",
"title": "1. Pixelwise Instance Segmentation with a Dynamically Instantiated Network",
"authors": "Anurag Arnab, Philip H. S. Torr",
"year": "2017"
},
{
"id": "1611.08303",
"title": "2. Deep Watershed Transform for Instance Segmentation",
"authors": "M. Bai, R. Urtasun",
"year": "2017"
},
{
"id": "1906.04016",
"title": "3. Learning Temporal Pose Estimation from Sparsely Labeled Videos",
"authors": "Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, Lorenzo Torresani",
"year": "2019"
},
{
"id": "1511.02674",
"title": "4. Semantic Segmentation with Boundary Neural Fields",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2016"
},
{
"id": "1803.05549",
"title": "5. Object Detection in Video with Spatiotemporal Sampling Networks",
"authors": "Gedas Bertasius, Lorenzo Torresani, Jianbo Shi",
"year": "2020"
}
] |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.