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 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
1912.04573 | 99 | 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 | 100 | 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 | 101 | 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 | 102 | 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 | 103 | 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 | 104 | 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 | 105 | 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 | 106 | 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 | 107 | 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 | 108 | 1/56g00yxKukElqbTfwUwxzalAwyaflXmZ5StmYDnnXUUXdkjCfnzsl504ZkFgbVwrJXP09kdPE2kkSuc6E4sguezPxP6+bYXwT5kKlGXLFFoviTBLUZPY7GQjDGcqJI5QZ4W4lbEQNZegSKrsQguWXV0nrshb4teDhqlq/LeIowSmcwQUEcA11uIcGNIHBGJ7hFd681Hvx3r2PReuaV8ycwB94nz+F64+t</latexit>3) | 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 | 109 | Propagated Instance Segmentation (time 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 | 110 | 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 | 111 | 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 | 112 | 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 | 113 | 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 | 114 | 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 | 115 | 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 | 116 | 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 | 117 | 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 | 118 | 2) Instance Feature Propagation (from time t to time 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 | 119 | 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 | 120 | 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 | 121 | 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 | 122 | 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 | 123 | 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 | 124 | 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 | 125 | 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 | 126 | ah+9UPFspgnyCQ1pue5Kfo51SiY5NNKPzM8pWxMh7xnaUJjbvx8fu2UnFklJJHSthIkc/X3RE5jYyZxYDtjiiOz7M3E/7xehtG1n4skzZAnbLEoyiRBRWavk1BozlBOLKFMC3srYSOqKUMbUMWG4C2/vEraF3XPrXv3l7XGTRFHGU7gFM7BgytowB00oQUMHuEZXuHNUc6L8+58LFpLTjFzDH/gfP4AkYyPHA==</latexit>1) | 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 | 127 | Instance Feature Computation (time t)Instancej | 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 | 128 | <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 | 129 | sha1_base64="ocTLVc574NOAe/DoMycp9A9e/3g=">AAAB6HicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cW7Ae0oWy2k3bbzSbsboQS+gu8eFDEqz/Jm//GbZuDtj4YeLw3w8y8IBFcG9f9dgobm1vbO8Xd0t7+weFR+fikpeNUMWyyWMSqE1CNgktsGm4EdhKFNAoEtoPJ/dxvP6HSPJaPZpqgH9Gh5CFn1FipMe6XK27VXYCsEy8nFchR75e/eoOYpRFKwwTVuuu5ifEzqgxnAmelXqoxoWxCh9i1V | 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 | 130 | oOYpRFKwwTVuuu5ifEzqgxnAmelXqoxoWxCh9i1VNIItZ8tDp2RC6sMSBgrW9KQhfp7IqOR1tMosJ0RNSO96s3F/7xuasJbP+MySQ1KtlwUpoKYmMy/JgOukBkxtYQyxe2thI2ooszYbEo2BG/15XXSuqp6btVrXFdqd3kcRTiDc7gED26gBg9QhyYwQHiGV3hzxs6L8+58LFsLTj5zCn/gfP4A0QGM7g==</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 | 131 | sha1_base64="ocTLVc574NOAe/DoMycp9A9e/3g=">AAAB6HicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cW7Ae0oWy2k3bbzSbsboQS+gu8eFDEqz/Jm//GbZuDtj4YeLw3w8y8IBFcG9f9dgobm1vbO8Xd0t7+weFR+fikpeNUMWyyWMSqE1CNgktsGm4EdhKFNAoEtoPJ/dxvP6HSPJaPZpqgH9Gh5CFn1FipMe6XK27VXYCsEy8nFchR75e/eoOYpRFKwwTVuuu5ifEzqgxnAmelXqoxoWxCh9i1V | 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 | 132 | oOYpRFKwwTVuuu5ifEzqgxnAmelXqoxoWxCh9i1VNIItZ8tDp2RC6sMSBgrW9KQhfp7IqOR1tMosJ0RNSO96s3F/7xuasJbP+MySQ1KtlwUpoKYmMy/JgOukBkxtYQyxe2thI2ooszYbEo2BG/15XXSuqp6btVrXFdqd3kcRTiDc7gED26gBg9QhyYwQHiGV3hzxs6L8+58LFsLTj5zCn/gfP4A0QGM7g==</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 | 133 | sha1_base64="ocTLVc574NOAe/DoMycp9A9e/3g=">AAAB6HicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cW7Ae0oWy2k3bbzSbsboQS+gu8eFDEqz/Jm//GbZuDtj4YeLw3w8y8IBFcG9f9dgobm1vbO8Xd0t7+weFR+fikpeNUMWyyWMSqE1CNgktsGm4EdhKFNAoEtoPJ/dxvP6HSPJaPZpqgH9Gh5CFn1FipMe6XK27VXYCsEy8nFchR75e/eoOYpRFKwwTVuuu5ifEzqgxnAmelXqoxoWxCh9i1V | 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 | 134 | oOYpRFKwwTVuuu5ifEzqgxnAmelXqoxoWxCh9i1VNIItZ8tDp2RC6sMSBgrW9KQhfp7IqOR1tMosJ0RNSO96s3F/7xuasJbP+MySQ1KtlwUpoKYmMy/JgOukBkxtYQyxe2thI2ooszYbEo2BG/15XXSuqp6btVrXFdqd3kcRTiDc7gED26gBg9QhyYwQHiGV3hzxs6L8+58LFsLTj5zCn/gfP4A0QGM7g==</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 | 135 | sha1_base64="ocTLVc574NOAe/DoMycp9A9e/3g=">AAAB6HicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cW7Ae0oWy2k3bbzSbsboQS+gu8eFDEqz/Jm//GbZuDtj4YeLw3w8y8IBFcG9f9dgobm1vbO8Xd0t7+weFR+fikpeNUMWyyWMSqE1CNgktsGm4EdhKFNAoEtoPJ/dxvP6HSPJaPZpqgH9Gh5CFn1FipMe6XK27VXYCsEy8nFchR75e/eoOYpRFKwwTVuuu5ifEzqgxnAmelXqoxoWxCh9i1V | 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 | 136 | oOYpRFKwwTVuuu5ifEzqgxnAmelXqoxoWxCh9i1VNIItZ8tDp2RC6sMSBgrW9KQhfp7IqOR1tMosJ0RNSO96s3F/7xuasJbP+MySQ1KtlwUpoKYmMy/JgOukBkxtYQyxe2thI2ooszYbEo2BG/15XXSuqp6btVrXFdqd3kcRTiDc7gED26gBg9QhyYwQHiGV3hzxs6L8+58LFsLTj5zCn/gfP4A0QGM7g==</latexit>Instancei<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 | 137 | sha1_base64="9jtkmSnVkzQh6f031uh1XGFh3yI=">AAAB6HicbVBNS8NAEJ3Ur1q/qh69LBbBU0lEqMeiF48t2A9oQ9lsJ+3azSbsboQS+gu8eFDEqz/Jm//GbZuDtj4YeLw3w8y8IBFcG9f9dgobm1vbO8Xd0t7+weFR+fikreNUMWyxWMSqG1CNgktsGW4EdhOFNAoEdoLJ3dzvPKHSPJYPZpqgH9GR5CFn1FipyQflilt1FyDrxMtJBXI0BuWv/jBmaYTSMEG17nluYvyMKsOZwFmpn2pMKJvQEfYsl | 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 | 138 | jBmaYTSMEG17nluYvyMKsOZwFmpn2pMKJvQEfYslTRC7WeLQ2fkwipDEsbKljRkof6eyGik9TQKbGdEzVivenPxP6+XmvDGz7hMUoOSLReFqSAmJvOvyZArZEZMLaFMcXsrYWOqKDM2m5INwVt9eZ20r6qeW/Wa15X6bR5HEc7gHC7BgxrU4R4a0AIGCM/wCm/Oo/PivDsfy9aCk8+cwh84nz/PfYzt</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 | 139 | sha1_base64="9jtkmSnVkzQh6f031uh1XGFh3yI=">AAAB6HicbVBNS8NAEJ3Ur1q/qh69LBbBU0lEqMeiF48t2A9oQ9lsJ+3azSbsboQS+gu8eFDEqz/Jm//GbZuDtj4YeLw3w8y8IBFcG9f9dgobm1vbO8Xd0t7+weFR+fikreNUMWyxWMSqG1CNgktsGW4EdhOFNAoEdoLJ3dzvPKHSPJYPZpqgH9GR5CFn1FipyQflilt1FyDrxMtJBXI0BuWv/jBmaYTSMEG17nluYvyMKsOZwFmpn2pMKJvQEfYsl | 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 | 140 | jBmaYTSMEG17nluYvyMKsOZwFmpn2pMKJvQEfYslTRC7WeLQ2fkwipDEsbKljRkof6eyGik9TQKbGdEzVivenPxP6+XmvDGz7hMUoOSLReFqSAmJvOvyZArZEZMLaFMcXsrYWOqKDM2m5INwVt9eZ20r6qeW/Wa15X6bR5HEc7gHC7BgxrU4R4a0AIGCM/wCm/Oo/PivDsfy9aCk8+cwh84nz/PfYzt</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 | 141 | sha1_base64="9jtkmSnVkzQh6f031uh1XGFh3yI=">AAAB6HicbVBNS8NAEJ3Ur1q/qh69LBbBU0lEqMeiF48t2A9oQ9lsJ+3azSbsboQS+gu8eFDEqz/Jm//GbZuDtj4YeLw3w8y8IBFcG9f9dgobm1vbO8Xd0t7+weFR+fikreNUMWyxWMSqG1CNgktsGW4EdhOFNAoEdoLJ3dzvPKHSPJYPZpqgH9GR5CFn1FipyQflilt1FyDrxMtJBXI0BuWv/jBmaYTSMEG17nluYvyMKsOZwFmpn2pMKJvQEfYsl | 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 | 142 | jBmaYTSMEG17nluYvyMKsOZwFmpn2pMKJvQEfYslTRC7WeLQ2fkwipDEsbKljRkof6eyGik9TQKbGdEzVivenPxP6+XmvDGz7hMUoOSLReFqSAmJvOvyZArZEZMLaFMcXsrYWOqKDM2m5INwVt9eZ20r6qeW/Wa15X6bR5HEc7gHC7BgxrU4R4a0AIGCM/wCm/Oo/PivDsfy9aCk8+cwh84nz/PfYzt</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 | 143 | sha1_base64="9jtkmSnVkzQh6f031uh1XGFh3yI=">AAAB6HicbVBNS8NAEJ3Ur1q/qh69LBbBU0lEqMeiF48t2A9oQ9lsJ+3azSbsboQS+gu8eFDEqz/Jm//GbZuDtj4YeLw3w8y8IBFcG9f9dgobm1vbO8Xd0t7+weFR+fikreNUMWyxWMSqG1CNgktsGW4EdhOFNAoEdoLJ3dzvPKHSPJYPZpqgH9GR5CFn1FipyQflilt1FyDrxMtJBXI0BuWv/jBmaYTSMEG17nluYvyMKsOZwFmpn2pMKJvQEfYsl | 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 | 144 | jBmaYTSMEG17nluYvyMKsOZwFmpn2pMKJvQEfYslTRC7WeLQ2fkwipDEsbKljRkof6eyGik9TQKbGdEzVivenPxP6+XmvDGz7hMUoOSLReFqSAmJvOvyZArZEZMLaFMcXsrYWOqKDM2m5INwVt9eZ20r6qeW/Wa15X6bR5HEc7gHC7BgxrU4R4a0AIGCM/wCm/Oo/PivDsfy9aCk8+cwh84nz/PfYzt</latexit>Instancej | 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 | 145 | <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 | 146 | sha1_base64="ocTLVc574NOAe/DoMycp9A9e/3g=">AAAB6HicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cW7Ae0oWy2k3bbzSbsboQS+gu8eFDEqz/Jm//GbZuDtj4YeLw3w8y8IBFcG9f9dgobm1vbO8Xd0t7+weFR+fikpeNUMWyyWMSqE1CNgktsGm4EdhKFNAoEtoPJ/dxvP6HSPJaPZpqgH9Gh5CFn1FipMe6XK27VXYCsEy8nFchR75e/eoOYpRFKwwTVuuu5ifEzqgxnAmelXqoxoWxCh9i1V | 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 | 147 | oOYpRFKwwTVuuu5ifEzqgxnAmelXqoxoWxCh9i1VNIItZ8tDp2RC6sMSBgrW9KQhfp7IqOR1tMosJ0RNSO96s3F/7xuasJbP+MySQ1KtlwUpoKYmMy/JgOukBkxtYQyxe2thI2ooszYbEo2BG/15XXSuqp6btVrXFdqd3kcRTiDc7gED26gBg9QhyYwQHiGV3hzxs6L8+58LFsLTj5zCn/gfP4A0QGM7g==</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 | 148 | sha1_base64="ocTLVc574NOAe/DoMycp9A9e/3g=">AAAB6HicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cW7Ae0oWy2k3bbzSbsboQS+gu8eFDEqz/Jm//GbZuDtj4YeLw3w8y8IBFcG9f9dgobm1vbO8Xd0t7+weFR+fikpeNUMWyyWMSqE1CNgktsGm4EdhKFNAoEtoPJ/dxvP6HSPJaPZpqgH9Gh5CFn1FipMe6XK27VXYCsEy8nFchR75e/eoOYpRFKwwTVuuu5ifEzqgxnAmelXqoxoWxCh9i1V | 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 | 149 | oOYpRFKwwTVuuu5ifEzqgxnAmelXqoxoWxCh9i1VNIItZ8tDp2RC6sMSBgrW9KQhfp7IqOR1tMosJ0RNSO96s3F/7xuasJbP+MySQ1KtlwUpoKYmMy/JgOukBkxtYQyxe2thI2ooszYbEo2BG/15XXSuqp6btVrXFdqd3kcRTiDc7gED26gBg9QhyYwQHiGV3hzxs6L8+58LFsLTj5zCn/gfP4A0QGM7g==</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 | 150 | sha1_base64="ocTLVc574NOAe/DoMycp9A9e/3g=">AAAB6HicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cW7Ae0oWy2k3bbzSbsboQS+gu8eFDEqz/Jm//GbZuDtj4YeLw3w8y8IBFcG9f9dgobm1vbO8Xd0t7+weFR+fikpeNUMWyyWMSqE1CNgktsGm4EdhKFNAoEtoPJ/dxvP6HSPJaPZpqgH9Gh5CFn1FipMe6XK27VXYCsEy8nFchR75e/eoOYpRFKwwTVuuu5ifEzqgxnAmelXqoxoWxCh9i1V | 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 | 151 | oOYpRFKwwTVuuu5ifEzqgxnAmelXqoxoWxCh9i1VNIItZ8tDp2RC6sMSBgrW9KQhfp7IqOR1tMosJ0RNSO96s3F/7xuasJbP+MySQ1KtlwUpoKYmMy/JgOukBkxtYQyxe2thI2ooszYbEo2BG/15XXSuqp6btVrXFdqd3kcRTiDc7gED26gBg9QhyYwQHiGV3hzxs6L8+58LFsLTj5zCn/gfP4A0QGM7g==</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 | 152 | sha1_base64="ocTLVc574NOAe/DoMycp9A9e/3g=">AAAB6HicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cW7Ae0oWy2k3bbzSbsboQS+gu8eFDEqz/Jm//GbZuDtj4YeLw3w8y8IBFcG9f9dgobm1vbO8Xd0t7+weFR+fikpeNUMWyyWMSqE1CNgktsGm4EdhKFNAoEtoPJ/dxvP6HSPJaPZpqgH9Gh5CFn1FipMe6XK27VXYCsEy8nFchR75e/eoOYpRFKwwTVuuu5ifEzqgxnAmelXqoxoWxCh9i1V | 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 | 153 | oOYpRFKwwTVuuu5ifEzqgxnAmelXqoxoWxCh9i1VNIItZ8tDp2RC6sMSBgrW9KQhfp7IqOR1tMosJ0RNSO96s3F/7xuasJbP+MySQ1KtlwUpoKYmMy/JgOukBkxtYQyxe2thI2ooszYbEo2BG/15XXSuqp6btVrXFdqd3kcRTiDc7gED26gBg9QhyYwQHiGV3hzxs6L8+58LFsLTj5zCn/gfP4A0QGM7g==</latexit>Instancei<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 | 154 | sha1_base64="9jtkmSnVkzQh6f031uh1XGFh3yI=">AAAB6HicbVBNS8NAEJ3Ur1q/qh69LBbBU0lEqMeiF48t2A9oQ9lsJ+3azSbsboQS+gu8eFDEqz/Jm//GbZuDtj4YeLw3w8y8IBFcG9f9dgobm1vbO8Xd0t7+weFR+fikreNUMWyxWMSqG1CNgktsGW4EdhOFNAoEdoLJ3dzvPKHSPJYPZpqgH9GR5CFn1FipyQflilt1FyDrxMtJBXI0BuWv/jBmaYTSMEG17nluYvyMKsOZwFmpn2pMKJvQEfYsl | 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 | 155 | jBmaYTSMEG17nluYvyMKsOZwFmpn2pMKJvQEfYslTRC7WeLQ2fkwipDEsbKljRkof6eyGik9TQKbGdEzVivenPxP6+XmvDGz7hMUoOSLReFqSAmJvOvyZArZEZMLaFMcXsrYWOqKDM2m5INwVt9eZ20r6qeW/Wa15X6bR5HEc7gHC7BgxrU4R4a0AIGCM/wCm/Oo/PivDsfy9aCk8+cwh84nz/PfYzt</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 | 156 | sha1_base64="9jtkmSnVkzQh6f031uh1XGFh3yI=">AAAB6HicbVBNS8NAEJ3Ur1q/qh69LBbBU0lEqMeiF48t2A9oQ9lsJ+3azSbsboQS+gu8eFDEqz/Jm//GbZuDtj4YeLw3w8y8IBFcG9f9dgobm1vbO8Xd0t7+weFR+fikreNUMWyxWMSqG1CNgktsGW4EdhOFNAoEdoLJ3dzvPKHSPJYPZpqgH9GR5CFn1FipyQflilt1FyDrxMtJBXI0BuWv/jBmaYTSMEG17nluYvyMKsOZwFmpn2pMKJvQEfYsl | 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 | 157 | jBmaYTSMEG17nluYvyMKsOZwFmpn2pMKJvQEfYslTRC7WeLQ2fkwipDEsbKljRkof6eyGik9TQKbGdEzVivenPxP6+XmvDGz7hMUoOSLReFqSAmJvOvyZArZEZMLaFMcXsrYWOqKDM2m5INwVt9eZ20r6qeW/Wa15X6bR5HEc7gHC7BgxrU4R4a0AIGCM/wCm/Oo/PivDsfy9aCk8+cwh84nz/PfYzt</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 | 158 | sha1_base64="9jtkmSnVkzQh6f031uh1XGFh3yI=">AAAB6HicbVBNS8NAEJ3Ur1q/qh69LBbBU0lEqMeiF48t2A9oQ9lsJ+3azSbsboQS+gu8eFDEqz/Jm//GbZuDtj4YeLw3w8y8IBFcG9f9dgobm1vbO8Xd0t7+weFR+fikreNUMWyxWMSqG1CNgktsGW4EdhOFNAoEdoLJ3dzvPKHSPJYPZpqgH9GR5CFn1FipyQflilt1FyDrxMtJBXI0BuWv/jBmaYTSMEG17nluYvyMKsOZwFmpn2pMKJvQEfYsl | 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 | 159 | jBmaYTSMEG17nluYvyMKsOZwFmpn2pMKJvQEfYslTRC7WeLQ2fkwipDEsbKljRkof6eyGik9TQKbGdEzVivenPxP6+XmvDGz7hMUoOSLReFqSAmJvOvyZArZEZMLaFMcXsrYWOqKDM2m5INwVt9eZ20r6qeW/Wa15X6bR5HEc7gHC7BgxrU4R4a0AIGCM/wCm/Oo/PivDsfy9aCk8+cwh84nz/PfYzt</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 | 160 | sha1_base64="9jtkmSnVkzQh6f031uh1XGFh3yI=">AAAB6HicbVBNS8NAEJ3Ur1q/qh69LBbBU0lEqMeiF48t2A9oQ9lsJ+3azSbsboQS+gu8eFDEqz/Jm//GbZuDtj4YeLw3w8y8IBFcG9f9dgobm1vbO8Xd0t7+weFR+fikreNUMWyxWMSqG1CNgktsGW4EdhOFNAoEdoLJ3dzvPKHSPJYPZpqgH9GR5CFn1FipyQflilt1FyDrxMtJBXI0BuWv/jBmaYTSMEG17nluYvyMKsOZwFmpn2pMKJvQEfYsl | 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 | 161 | jBmaYTSMEG17nluYvyMKsOZwFmpn2pMKJvQEfYslTRC7WeLQ2fkwipDEsbKljRkof6eyGik9TQKbGdEzVivenPxP6+XmvDGz7hMUoOSLReFqSAmJvOvyZArZEZMLaFMcXsrYWOqKDM2m5INwVt9eZ20r6qeW/Wa15X6bR5HEc7gHC7BgxrU4R4a0AIGCM/wCm/Oo/PivDsfy9aCk8+cwh84nz/PfYzt</latexit>Figure | 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 | 162 | 4: An illustration of the 3 steps implemented by our mask propagation branch: 1)For every detected instance in frame | 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 | 163 | t, we compute an instance-specific feature tensor via element-wise multiplication between tensor tand the given frame-level
instance mask at frame t.2)Next, the element-wise difference of the feature tensors associated with frames tandt+is
used to predict motion offsets between frames tandt+. The resulting offsets are used to propagate the instance-specific
tensors from time tto timet+via deformable convolution. The propagated tensors represent instance-specific features
predicted for time t+using the tensors computed at time t.3)Lastly, we add the propagated instance feature tensors to the
tensor effectively computed at t+. A convolutional layer applied to these tensors predicts instance masks in frame t+.
The pixels that do not belong to any object instances are zeroed out using an instance-agnostic attention map.
the tensorft+effectively computed from the frame at time
t+. If the object instance prediction is consistent with the
feature computation, the feature tensors will be aligned and
thus, they will reinforce each other in the predicted region.
Finally, the resulting feature tensors i
t;t+are fed into
a11convolution layer that outputs instance masks for | 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 | 164 | Finally, the resulting feature tensors i
t;t+are fed into
a11convolution layer that outputs instance masks for
each object instance iin framet+. The masks are normalized with the softmax nonlinearity across all Ntinstances. To zero-out pixels that do not belong to any object
instance, we use a single 33convolution that computes
an instance-agnostic attention map At+from feature tensorft+. We then multiply At+with each of our predicted
instance masks. A detailed illustration of our mask propagation branch is presented in Figure 4.
4.3. Video-Level Segmentation Instances
Given a video of length L, our goal is to produce videolevel segmentation instances Mi2 RLHW. Conceptually, this requires linking clip-level instance tracks
Mi
t T:t+TandMj
t0 T:t0+Twheniandjrepresent the same
object instance, i.e., when the instances are matching. We
achieve this by assigning a video-level instance ID to each
of our predicted clip-level instance tracks. Matching instance tracks are assigned the same video-level instance ID.
Matching Clip-Level Instance Tracks. Consider a pair of
clip-level instance tracks 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 | 165 | Matching Clip-Level Instance Tracks. Consider a pair of
clip-level instance tracks Mi
t T:t+TandMj
t0 T:t0+T, that
are centered around frames tandt0respectively. They overlap in time ifjt t0j<2T+ 1. Let us denote their overlapping time interval as \t;t0. Given two overlapping in-stance tracks, we can check if they match by comparing
their predicted instance masks in the overlapping frames.
If the masks of the two instance tracks align well, then they
are likely to encode the same object instance. Otherwise,
each track represents a different object instance. We compute a matching score mi;j
t;t0between two clip-level instance
tracks using our previously defined soft IoU metric as:
mi;j
t;t0=1
j\t;t0jX
~t2\t;t0sIoU (Mi
t T:t+T(~t); Mj
t0 T:t0+T(~t))(4)
Video-Level Instance ID Assignment. We denote withY
the set of video-level instance IDs. The set Yis built incrementally by matching clip-level instance tracks Mi
t T:t+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 | 166 | the set of video-level instance IDs. The set Yis built incrementally by matching clip-level instance tracks Mi
t T:t+T
in order from time t= 1 to timet=L. Initially, we set
Y=f1;2;:::;N 1gwhereN1is the number of object instances detected at time t= 1 (i.e., in the first clip of the
video). Let yi
t2Y denote the video-level instance ID assigned to clip-level instance track Mi
t T:t+T. As we move
forward in time t>1, the video-level instance ID yi
tis assigned by matching clip-level instance Mi
t T:t+Tto all previously processed instance tracks Mj
t0 T:t0+Tthat overlap
with this clip, i.e., such that \t;t06=;. For each video-level
instance IDy2Y already in the ID set, we compute a score
qi
t(y)capturing how well Mi
t T:t+Tmatches the tracks that
have been already assigned video-level ID y:
qi
t(y) =P
t0s.t.\t;t06=;PNt0
j=11fyj | 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 | 167 | qi
t(y) =P
t0s.t.\t;t06=;PNt0
j=11fyj
t0=ygmi;j
t;t0
P
t0s.t.\t;t06=;PNt0
j=11fyj
t0=yg(5)
time ttime t+5time t+10time t+15time t-15time t-10time t-5
Figure 5: An illustration of instance-specific features propagated from frame tto other frames in the given video clip. Here,
we visualize propagated activations from one randomly selected feature channel. The activations in the two rows correspond
to two different object instances detected at time t. Our visualizations suggest that MaskProp reliably propagates features
that are specific to each instance even when instances appear next to each other, and despite the changes in shape, pose and
the nuisances effects of deformation and occlusion.
where 1fyj
t0=ygis an indicator function that is equal
to1whenyj
t0=y, and 0otherwise.Nt0is the number
of detected instances at time t0, andmi;j
t;t0is the previously
defined matching score between instance tracks 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 | 168 | of detected instances at time t0, andmi;j
t;t0is the previously
defined matching score between instance tracks Mi
t T:t+T
andMj
t0 T:t0+T. The score qi
t(y)is effectively an average
of the soft IoU computed between instance track Mi
t T:t+T,
and all its overlapping instance tracks that have been assigned video-level instance ID y.
Letq= maxy2Yqi
t(y)be the maximum score obtained
by considering all possible video-level instance IDs y2Y.
Ifqis greater than a certain threshold, the current instance track Mi
t T:t+Tis assigned the video-level instance
IDy= arg maxy2Yqi
t(y). Otherwise, the clip-level track
does not match any of the current video-level instances. In
such case we create a new video-level instance ID and assign it to the the clip-level track while also expanding the
setY, i.e.,yi
t=jYj+ 1andY=Y[fjYj + 1g.
Finally, for each video-level instance ID y2 Y , we
generate the final sequence of segmentation instance masks | 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 | 169 | Finally, for each video-level instance ID y2 Y , we
generate the final sequence of segmentation instance masks
My2RLHWas:
My(t) =(
Mi
t T:t+T(t)ifyi
t=y
0 otherwise .4.4. Implementation Details
Backbone Network. As our backbone we use a Spatiotemporal Sampling Network [5] based on a Deformable
ResNeXt-101-64x4d [12, 41] with a feature pyramid network (FPN) [24] attached on top of it.
Detection Network. For detection, we use a Hybrid Task
Cascade Network [7] with a 3stage cascade.
Mask Propagation Branch. The residual block in the
mask propagation branch consists of two 33convolutions with 128output channels each. The instance feature
propagation is applied to the FPN feature map with the second largest spatial resolution. To propagate instance features from one frame to another, we predict 9 (x;y)offsets
for every pixel, which are then used as input to a 33deformable convolutional layer with 256channels. To capture
motion at different scales, we use three levels of dilated deformable convolutions with dilation rates 3;6;12as in [3]. | 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 | 170 | motion at different scales, we use three levels of dilated deformable convolutions with dilation rates 3;6;12as in [3].
High-Resolution Mask Refinement. Predicting masks
from RoI features typically leads to low resolution predictions. We address this issue via a high-resolution mask refinement step. Given the center of a detected bounding box,
we crop a 384384patch around the object, preserving
the original aspect ratio. We then feed the RGB patch and
the predicted low-resolution mask through 3residual blocks
each with 128channels to obtain a high-resolution mask.
Method Pre-training Data mAP AP@75 AR@1 AR@10
DeepSORTz[39] Imagenet [34], COCO [25] 26.1 26.1 27.8 31.3
FEELVOSz[37] Imagenet [34], COCO [25] 26.9 29.7 29.9 33.4
OSMNz[43] Imagenet [34], COCO [25] 27.5 29.1 28.6 33.1
MaskTrack R-CNNz[42] Imagenet [34], COCO [25] 30.3 32.6 31.0 35.5 | 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 | 171 | MaskTrack R-CNNz[42] Imagenet [34], COCO [25] 30.3 32.6 31.0 35.5
MaskTrack R-CNNImagenet [34], COCO [25] 36.9 40.2 34.3 42.9
EnsembleVIS [28] Imagenet [34], COCO [25], Instagram [30], OpenImages [23] 44.8 48.9 42.7 51.7
MaskProp Imagenet [34], COCO [25] 46.6 51.2 44.0 52.6
MaskProp Imagenet [34], COCO [25], OpenImages [23] 50.0 55.9 44.6 54.5
Table 2: The results of video instance segmentation on the YouTube-VIS [42] validation dataset. We evaluate the performance
of each method according to mean average precision (mAP), average precision at 75% IoU threshold (AP@75), and average
recall given top 1(AR@1) and top 10(AR@10) detections. The baselines denoted withzwere implemented by the authors
in [42], whereas the methods marked withwere implemented by us, and use the same backbone and detection networks as
our approach. Despite its simplicity, MaskProp outperforms all prior video instance segmentation methods by a large margin. | 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 | 172 | our approach. Despite its simplicity, MaskProp outperforms all prior video instance segmentation methods by a large margin.
Scoring Video-Level Sequences. Each video-level sequence contains a list of classification scores and predicted
object labels. To assign a confidence score to each videolevel sequence, we average classification scores associated
with that sequence (separately for each object category).
Training. We use a similar training setup as in [7]. The loss
weights for each of the three cascade stages are set to 1;0:5;
and0:25respectively. The loss weight for the semantic segmentation branch is set to 0:1. We train our model on pairs
of frames, where the second frame in a pair is randomly selected with a time gap 2[ 25;25]relative to first frame.
We use a multi-scale training approach implemented by resizing the shorter side of the frame randomly between 400
and800pixels. Our model is trained in a distributed setting using 64GPUs, each GPU holding a single clip. The
training is done for 20epochs with an initial learning rate
of0:008, which is decreased by 10at16and19epochs.
We initialize our model with a Mask R-CNN pretrained on | 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 | 173 | of0:008, which is decreased by 10at16and19epochs.
We initialize our model with a Mask R-CNN pretrained on
COCO for the instance segmentation. The hyperparameters
of RPN and FPN are the same as in [7].
Inference. During testing, we run the bounding box prediction branch on 1000 proposals, apply non-maximum suppression, and use boxes with a score higher than 0:1as input to the mask prediction and mask propagation branches.
During inference, our MaskProp is applied to video clips
consisting of 13frames.
5. Experimental Results
In this section, we evaluate MaskProp for video instance
segmentation on YouTube-VIS [42], which contains 2;238
training, 302validation, and 343test videos. Each video
is annotated with per-pixel segmentation, category, and instance labels. The dataset contains 40object categories.
Since the evaluation on the test set is currently closed, we
perform our evaluations on the validation set.
5.1. Quantitative Results
Video instance segmentation is a very recent task [42],
and thus, there are only a few established baselines that we
7 9 11 13 15 17 19 21 | 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 | 174 | Video instance segmentation is a very recent task [42],
and thus, there are only a few established baselines that we
7 9 11 13 15 17 19 21
Instance Track Clip Length43.54444.54545.54646.547Accuracy (Video mAP)Figure 6: We plot video mAP as a function of an instance
track clip length (denoted as 2T+ 1in the paper). Based on
these results, we observe that optimal video instance segmentation performance is achieved when we propagate instance masks to T= 6previous and subsequent frames.
can compare our work to. We include in our comparison
MaskTrack R-CNN [42], and the EnsembleVIS method,
which won the ICCV 2019 video instance segmentation
challenge [28]. Additionally, to make the comparison with
MaskTrack R-CNN more fair, we reimplement it using the
same backbone and detection networks as our MaskProp
(see MaskTrack R-CNN* in Table 2).
We present our quantitative results in Table 2, where we
assess each method according to 1) mean video average precision (mAP), 2) video average precision at IoU threshold
of75%, and 3) average recall given 1and10highest scored
instances per video. From these results, we observe that
our MaskProp outperforms all the other baselines according | 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 | 175 | of75%, and 3) average recall given 1and10highest scored
instances per video. From these results, we observe that
our MaskProp outperforms all the other baselines according
to all four evaluation metrics, thus achieving state-of-theart results in video instance segmentation on YouTube-VIS.
It can be noted that we outperform EnsembleVIS [28] by
1:8mAP even though our approach is much simpler and
even if it uses orders of magnitude less labeled data for
pre-training. Furthermore, compared to the MaskTrack RCNN, our method achieves a 16:3improvement in mAP. We
also note that our implementation of MaskTrack [42] significantly improves upon the original work, but it is still 9:7
Method mAP AP@75
FlowNet2 Propagation 31.4 33.6
MaskTrack R-CNN36.9 40.2
MaskProp 46.6 51.2
Table 3: Here, we study the effectiveness of our mask propagation branch. If we replace it with the FlowNet2 propagation scheme, where masks are propagated using the optical
flow predicted by a FlowNet2 network [20], the accuracy
drops from 46:6mAP to 31:4mAP. Similarly, if we replace
our mask propagation branch with the tracking branch from | 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 | 176 | drops from 46:6mAP to 31:4mAP. Similarly, if we replace
our mask propagation branch with the tracking branch from
MaskTrack R-CNN, the accuracy drops to 36:9mAP. Note
that all of these baselines are implemented using the same
backbone and detection networks.
worse in mAP compared to our model. Lastly, we note that
additionally pre-training MaskProp on the OpenImages [23]
dataset as was done in [28] further boosts our performance
to50:0mAP.
5.2. Ablation Experiments
Mask Propagation Branch. To investigate the effectiveness of our mask propagation branch, we compare our
method with a FlowNet2 [20] propagation baseline. For
this baseline, we use exactly the same setup as for our
MaskProp, except that instance masks are propagated using the optical flow predicted by a FlowNet2 network [20]
rather than our proposed mask propagation scheme. For a
more complete comparison, we also include the MaskTrack
R-CNNfrom Table 2, which uses the originally proposed
tracking branch [42], but is implemented using the same
backbone and detection networks as our MaskProp.
These baselines allow us to directly compare the effectiveness of our mask propagation scheme versus the propagation mechanisms employed by FlowNet2 and MaskTrack | 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 | 177 | backbone and detection networks as our MaskProp.
These baselines allow us to directly compare the effectiveness of our mask propagation scheme versus the propagation mechanisms employed by FlowNet2 and MaskTrack
R-CNN [42] methods. The results in Table 3 show that
MaskProp outperforms these baselines by a large margin.
Instance Track Clip Length. Due to occlusions, object instances in video may not be visible in some frames. If there
are occlusions separated by 2T0+ 2time-steps, we can use
T >T0to predict longer clip-level instance tracks. In Figure 6, we study video instance segmentation performance as
a function of instance track clip length (denoted as 2T+ 1
in the paper). Our results indicate that the best accuracy is
achieved when we use a clip length of 13, meaning that we
propagate instances to 6previous and 6subsequent frames.
High-Resolution Mask Refinement. We also study the impact of our high-resolution mask refinement, described in
Subsection 4.4. We report that removing this refinement
causes a drop of 1:9%in video instance segmentation mAP.
Importance of Frame-Level Instance Masks. As described in our main draft, we use frame-level instance masks | 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 | 178 | Importance of Frame-Level Instance Masks. As described in our main draft, we use frame-level instance masks
for instance-specific feature computation. To investigate theBackbone Network mAP AP@75
ResNet-50 40.0 42.9
ResNet-101 42.5 45.6
ResNeXt-101-64x4d 44.3 48.3
STSN [5]-ResNeXt-101-64x4d 46.6 51.2
Table 4: We study the effect of frame-level instance masks
to our system’s performance. We evaluate our method’s
accuracy when using instance masks obtained from Mask
R-CNN with several different backbones. Our results indicate that frame-level instance masks obtained from stronger
models lead to better video instance segmentation results.
contribution of these masks to the performance of our system, we experiment with masks obtained from several different Mask R-CNN models. In Table 4, we present our
results for this ablation. Our results indicate that framelevel instance masks obtained from stronger models allow
us to achieve better video instance segmentation performance. Thus, we expect that future improvements in image
instance segmentation will further benefit our method.
5.3. Qualitative Results
In Figure 7, we compare our predicted clip-level instance | 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 | 179 | instance segmentation will further benefit our method.
5.3. Qualitative Results
In Figure 7, we compare our predicted clip-level instance
tracks (last row of predictions for each clip) with the MaskTrack R-CNN predictions (first row of predictions). We
use different colors to represent different object instances.
Our qualitative results suggest that our MaskProp produces
more robust and temporally coherent instance tracks than
MaskTrack R-CNN. Such differences in performance are
especially noticeable when a video contains large object
motion, occlusions, or overlapping objects.
In Figure 5, we also visualize instance-specific features
that are propagated from frame tto other frames in the
given video clip for two different object instances detected
in framet. Here, we show activations from a randomly selected feature channel. Based on these results, we observe
that our MaskProp reliably propagates features that are specific to each instance despite motion blur, object deformations and large variations in object appearance.
6. Conclusion
In this work, we introduced MaskProp, a novel architecture for video instance segmentation. Our method is conceptually simple, it does not require large amounts of labeled data for pre-training, and it produces state-of-the-art | 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 | 180 | results on YouTube-VIS dataset. In future, we plan to extend MaskProp to scenarios where only bounding box annotations are available. We are also interested in applying our
method to problems such as pose estimation and tracking.
Video ClipMaskPropMaskTrack R-CNN
Video ClipMaskPropMaskTrack R-CNN
Video Clip
MaskProp
MaskTrack R-CNNVideo Clip
MaskPropMaskTrack R-CNN
Figure 7: We compare our video instance segmentation results with MaskTrack R-CNN [42] predictions. Different object
instances are encoded with different colors. The first row for each video shows the original frames. The second row illustrates
the mask predictions of MaskTrack R-CNN and the third row those obtained with our MaskProp. Compared to MaskTrack
R-CNN, our MaskProp tracks object instances more robustly even when they are occluded or overlap with each other.
References
[1] Anurag Arnab and Philip H. S. Torr. Pixelwise instance segmentation with a dynamically instantiated network. In 2017
IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017 ,
pages 879–888, 2017. 2
[2] M. Bai and R. Urtasun. Deep watershed transform for instance segmentation. In 2017 IEEE Conference on Computer | 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 | 181 | pages 879–888, 2017. 2
[2] M. Bai and R. Urtasun. Deep watershed transform for instance segmentation. In 2017 IEEE Conference on Computer
Vision and Pattern Recognition (CVPR) , pages 2858–2866,
July 2017. 2
[3] Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo
Shi, and Lorenzo Torresani. Learning temporal pose estimation from sparsely labeled videos. In Advances in Neural
Information Processing Systems 33 . 2019. 4, 6
[4] Gedas Bertasius, Jianbo Shi, and Lorenzo Torresani. Semantic segmentation with boundary neural fields. In The
IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , June 2016. 1
[5] Gedas Bertasius, Lorenzo Torresani, and Jianbo Shi. Object
detection in video with spatiotemporal sampling networks.
InECCV (12) , volume 11216 of Lecture Notes in Computer
Science , pages 342–357. Springer, 2018. 2, 4, 6, 8
[6] Sergi Caelles, Kevis-Kokitsi Maninis, Jordi Pont-Tuset, | 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 | 182 | [6] Sergi Caelles, Kevis-Kokitsi Maninis, Jordi Pont-Tuset,
Laura Leal-Taixé, Daniel Cremers, and Luc Van Gool. Oneshot video object segmentation. In Computer Vision and Pattern Recognition (CVPR) , 2017. 2
[7] Kai Chen, Jiangmiao Pang, Jiaqi Wang, Yu Xiong, Xiaoxiao
Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jianping Shi,
Wanli Ouyang, Chen Change Loy, and Dahua Lin. Hybrid
task cascade for instance segmentation. In IEEE Conference
on Computer Vision and Pattern Recognition, CVPR 2019,
Long Beach, CA, USA, June 16-20, 2019 , pages 4974–4983.
Computer Vision Foundation / IEEE, 2019. 6, 7
[8] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos,
Kevin Murphy, and Alan L. Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE Trans. Pattern Anal.
Mach. Intell. , 40(4):834–848, 2018. 1 | 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 | 183 | Mach. Intell. , 40(4):834–848, 2018. 1
[9] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian
Schroff, and Hartwig Adam. Encoder-decoder with atrous
separable convolution for semantic image segmentation. In
Computer Vision - ECCV 2018 - 15th European Conference,
Munich, Germany, September 8-14, 2018, Proceedings, Part
VII, pages 833–851, 2018. 2
[10] Yuhua Chen, Jordi Pont-Tuset, Alberto Montes, and Luc Van
Gool. Blazingly fast video object segmentation with pixelwise metric learning. In CVPR , pages 1189–1198. IEEE
Computer Society, 2018. 2
[11] Jifeng Dai, Kaiming He, and Jian Sun. Instance-aware semantic segmentation via multi-task network cascades. In
CVPR , 2016. 2
[12] J. Dai, H. Qi, Y . Xiong, Y . Li, G. Zhang, H. Hu, and Y . Wei.
Deformable convolutional networks. In 2017 IEEE International Conference on Computer Vision (ICCV) , volume 00, | 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 | 184 | Deformable convolutional networks. In 2017 IEEE International Conference on Computer Vision (ICCV) , volume 00,
pages 764–773, Oct. 2017. 4, 6[13] Christoph Feichtenhofer, Axel Pinz, and Andrew Zisserman.
Detect to track and track to detect. In International Conference on Computer Vision (ICCV) , 2017. 2
[14] Ross Girshick. Fast R-CNN. In Proceedings of the International Conference on Computer Vision (ICCV) , 2015. 1
[15] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra
Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition
(CVPR) , 2014. 1
[16] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask R-CNN. In Proceedings of the International
Conference on Computer Vision (ICCV) , 2017. 1, 2, 4
[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun.
Deep residual learning for image recognition. 2016 IEEE
Conference on Computer Vision and Pattern Recognition
(CVPR) , pages 770–778, 2016. 1, 4 | 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 | 185 | Deep residual learning for image recognition. 2016 IEEE
Conference on Computer Vision and Pattern Recognition
(CVPR) , pages 770–778, 2016. 1, 4
[18] Alexander Hermans*, Lucas Beyer*, and Bastian Leibe. In
Defense of the Triplet Loss for Person Re-Identification.
arXiv preprint arXiv:1703.07737 , 2017. 2
[19] Alba Herrera-Palacio, Carles Ventura, and Xavier Giro-i Nieto. Video object linguistic grounding. In 1st International
Workshop on Multimodal Understanding and Learning for
Embodied Applications , MULEA ’19, pages 49–51, New
York, NY , USA, 2019. ACM. 2
[20] Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper,
Alexey Dosovitskiy, and Thomas Brox. Flownet 2.0: Evolution of optical flow estimation with deep networks. CoRR ,
abs/1612.01925, 2016. 8
[21] Suyog Dutt Jain, Bo Xiong, and Kristen Grauman. Fusionseg: Learning to combine motion and appearance for fully
automatic segmentation of generic objects in videos. In 2017 | 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 | 186 | automatic segmentation of generic objects in videos. In 2017
IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017 ,
pages 2117–2126, 2017. 2
[22] Alexander Kirillov, Evgeny Levinkov, Bjoern Andres, Bogdan Savchynskyy, and Carsten Rother. InstanceCut: from
edges to instances with multicut. In CVPR , 2017. 2
[23] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan
Popov, Matteo Malloci, Tom Duerig, and Vittorio Ferrari.
The open images dataset v4: Unified image classification,
object detection, and visual relationship detection at scale,
2018. 2, 7, 8
[24] Tsung-Yi Lin, Piotr Dollár, Ross Girshick, Kaiming He,
Bharath Hariharan, and Serge Belongie. Feature pyramid
networks for object detection. In CVPR , 2017. 6
[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, | 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 | 187 | networks for object detection. In CVPR , 2017. 6
[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays,
Pietro Perona, Deva Ramanan, Piotr Dollár, and C. Lawrence
Zitnick. Microsoft coco: Common objects in context. In
European Conference on Computer Vision (ECCV) , Zürich,
September 2014. 2, 3, 7
[26] S. Liu, J. Jia, S. Fidler, and R. Urtasun. Sgn: Sequential
grouping networks for instance segmentation. In 2017 IEEE
International Conference on Computer Vision (ICCV) , pages
3516–3524, Oct 2017. 2
[27] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully
convolutional networks for semantic segmentation. In
CVPR , 2015. 1
[28] Jonathon Luiten, Philip H.S. Torr, and Bastian Leibe. Video
instance segmentation 2019: A winning approach for combined detection, segmentation, classification and tracking.
In2019 IEEE International Conference on Computer Vision
Workshops, ICCV Workshops 2019, Seoul, South Korea, October 27-November 2, 2019 . IEEE, 2019. 1, 2, 7, 8 | 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 | 188 | Workshops, ICCV Workshops 2019, Seoul, South Korea, October 27-November 2, 2019 . IEEE, 2019. 1, 2, 7, 8
[29] Jonathon Luiten, Paul V oigtlaender, and Bastian Leibe. Premvos: Proposal-generation, refinement and merging for
video object segmentation. In Asian Conference on Computer Vision , 2018. 2
[30] Dhruv Mahajan, Ross Girshick, Vignesh Ramanathan,
Kaiming He, Manohar Paluri, Yixuan Li, Ashwin Bharambe,
and Laurens van der Maaten. Exploring the limits of weakly
supervised pretraining. In ECCV , 2018. 2, 7
[31] Aljoša Ošep, Paul V oigtlaender, Jonathon Luiten, Stefan
Breuers, and Bastian Leibe. Large-scale object mining for
object discovery from unlabeled video. ICRA , 2019. 2
[32] F. Perazzi, A. Khoreva, R. Benenson, B. Schiele, and
A.Sorkine-Hornung. Learning video object segmentation
from static images. In Computer Vision and Pattern Recognition , 2017. 2 | 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 | 189 | A.Sorkine-Hornung. Learning video object segmentation
from static images. In Computer Vision and Pattern Recognition , 2017. 2
[33] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun.
Faster R-CNN: Towards real-time object detection with region proposal networks. In Neural Information Processing
Systems (NIPS) , 2015. 1
[34] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy,
Aditya Khosla, Michael Bernstein, Alexander C. Berg, and
Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. arXiv:1409.0575 , 2014. 2, 7
[35] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz.
PWC-Net: CNNs for optical flow using pyramid, warping,
and cost volume. 2018. 2
[36] P. Tokmakov, K. Alahari, and C. Schmid. Learning video
object segmentation with visual memory. In 2017 IEEE International Conference on Computer Vision (ICCV) , pages
4491–4500, Oct 2017. 2 | 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 | 190 | object segmentation with visual memory. In 2017 IEEE International Conference on Computer Vision (ICCV) , pages
4491–4500, Oct 2017. 2
[37] Paul V oigtlaender, Yuning Chai, Florian Schroff, Hartwig
Adam, Bastian Leibe, and Liang-Chieh Chen. Feelvos: Fast
end-to-end embedding learning for video object segmentation. In CVPR , 2019. 2, 7
[38] Paul V oigtlaender, Michael Krause, Aljosa Osep, Jonathon
Luiten, Berin Balachandar Gnana Sekar, Andreas Geiger,
and Bastian Leibe. Mots: Multi-object tracking and segmentation. In Proceedings IEEE Conf. on Computer Vision and
Pattern Recognition (CVPR) , June 2019. 2
[39] Nicolai Wojke, Alex Bewley, and Dietrich Paulus. Simple
online and realtime tracking with a deep association metric.
CoRR , abs/1703.07402, 2017. 7
[40] Fanyi Xiao and Yong Jae Lee. Video object detection with an
aligned spatial-temporal memory. In European Conference
on Computer Vision (ECCV) , 2018. 2 | 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 | 191 | [40] Fanyi Xiao and Yong Jae Lee. Video object detection with an
aligned spatial-temporal memory. In European Conference
on Computer Vision (ECCV) , 2018. 2
[41] Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, and
Kaiming He. Aggregated residual transformations for deep
neural networks. In CVPR , 2017. 6[42] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In ICCV , 2019. 1, 2, 3, 7, 8, 9
[43] Linjie Yang, Yanran Wang, Xuehan Xiong, Jianchao Yang,
and Aggelos K. Katsaggelos. Efficient video object segmentation via network modulation. In 2018 IEEE Conference on
Computer Vision and Pattern Recognition, CVPR 2018, Salt
Lake City, UT, USA, June 18-22, 2018 , pages 6499–6507,
2018. 2, 7
[44] Jifeng Dai Xiangyang Ji Yi Li, Haozhi Qi and Yichen Wei.
Fully convolutional instance-aware semantic segmentation.
InCVPR , 2017. 2
[45] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang | 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 | 192 | InCVPR , 2017. 2
[45] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang
Wang, and Jiaya Jia. Pyramid scene parsing network. In 2017
IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017 ,
pages 6230–6239, 2017. 1
[46] Xizhou Zhu, Yujie Wang, Jifeng Dai, Lu Yuan, and Yichen
Wei. Flow-guided feature aggregation for video object detection. In International Conference on Computer Vision
(ICCV) , 2017. 2
[47] Idil Esen Zulfikar, Jonathon Luiten, and Bastian Leibe. Unovost: Unsupervised offline video object segmentation and
tracking for the 2019 unsupervised davis challenge. 2 | 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"
}
] |
2005.00928 | 0 | Quantifying Attention Flow in Transformers
Samira Abnar
ILLC, University of Amsterdam
[email protected] Zuidema
ILLC, University of Amsterdam
[email protected]
Abstract
In the Transformer model, “self-attention”
combines information from attended embeddings into the representation of the focal embedding in the next layer. Thus, across layers of the Transformer, information originating
from different tokens gets increasingly mixed.
This makes attention weights unreliable as explanations probes. In this paper, we consider
the problem of quantifying this flow of information through self-attention. We propose two
methods for approximating the attention to input tokens given attention weights, attention
rollout andattention flow , as post hoc methods
when we use attention weights as the relative
relevance of the input tokens. We show that
these methods give complementary views on
the flow of information, and compared to raw
attention, both yield higher correlations with
importance scores of input tokens obtained using an ablation method and input gradients.
1 Introduction
Attention (Bahdanau et al., 2015; Vaswani et al.,
2017) has become the key building block of neural sequence processing models, and visualizing
attention weights is the easiest and most popular | 2005.00928 | Quantifying Attention Flow in Transformers | In the Transformer model, "self-attention" combines information from attended
embeddings into the representation of the focal embedding in the next layer.
Thus, across layers of the Transformer, information originating from different
tokens gets increasingly mixed. This makes attention weights unreliable as
explanations probes. In this paper, we consider the problem of quantifying this
flow of information through self-attention. We propose two methods for
approximating the attention to input tokens given attention weights, attention
rollout and attention flow, as post hoc methods when we use attention weights
as the relative relevance of the input tokens. We show that these methods give
complementary views on the flow of information, and compared to raw attention,
both yield higher correlations with importance scores of input tokens obtained
using an ablation method and input gradients. | http://arxiv.org/pdf/2005.00928 | [
"Samira Abnar",
"Willem Zuidema"
] | [
"cs.LG",
"cs.AI",
"cs.CL"
] | null | null | cs.LG | 20200502 | 20200531 | [] |
2005.00928 | 1 | 2017) has become the key building block of neural sequence processing models, and visualizing
attention weights is the easiest and most popular
approach to interpret a model’s decisions and to
gain insights about its internals (Vaswani et al.,
2017; Xu et al., 2015; Wang et al., 2016; Lee et al.,
2017; Dehghani et al., 2019; Rockt ¨aschel et al.,
2016; Chen and Ji, 2019; Coenen et al., 2019; Clark
et al., 2019). Although it is wrong to equate attention with explanation (Pruthi et al., 2019; Jain and
Wallace, 2019), it can offer plausible and meaningful interpretations (Wiegreffe and Pinter, 2019;
Vashishth et al., 2019; Vig, 2019). In this paper,
we focus on problems arising when we move to the
higher layers of a model, due to lack of token identifiability of the embeddings in higher layers (Brunner et al., 2020).We propose two simple but effective methods to
compute attention scores to input tokens (i.e., token
attention ) at each layer, by taking raw attentions
(i.e., embedding attention ) of that layer as well as
those from the precedent layers. These methods | 2005.00928 | Quantifying Attention Flow in Transformers | In the Transformer model, "self-attention" combines information from attended
embeddings into the representation of the focal embedding in the next layer.
Thus, across layers of the Transformer, information originating from different
tokens gets increasingly mixed. This makes attention weights unreliable as
explanations probes. In this paper, we consider the problem of quantifying this
flow of information through self-attention. We propose two methods for
approximating the attention to input tokens given attention weights, attention
rollout and attention flow, as post hoc methods when we use attention weights
as the relative relevance of the input tokens. We show that these methods give
complementary views on the flow of information, and compared to raw attention,
both yield higher correlations with importance scores of input tokens obtained
using an ablation method and input gradients. | http://arxiv.org/pdf/2005.00928 | [
"Samira Abnar",
"Willem Zuidema"
] | [
"cs.LG",
"cs.AI",
"cs.CL"
] | null | null | cs.LG | 20200502 | 20200531 | [] |
2005.00928 | 2 | attention ) at each layer, by taking raw attentions
(i.e., embedding attention ) of that layer as well as
those from the precedent layers. These methods
are based on modelling the information flow in the
network with a DAG (Directed Acyclic Graph), in
which the nodes are input tokens and hidden embeddings, edges are the attentions from the nodes
in each layer to those in the previous layer, and
the weights of the edges are the attention weights.
The first method, attention rollout , assumes that
the identities of input tokens are linearly combined
through the layers based on the attention weights.
To adjust attention weights, it rolls out the weights
to capture the propagation of information from input tokens to intermediate hidden embeddings. The
second method, attention flow , considers the attention graph as a flow network. Using a maximum
flow algorithm, it computes maximum flow values,
from hidden embeddings (sources) to input tokens
(sinks). In both methods, we take the residual connection in the network into account to better model
the connections between input tokens and hidden
embedding. We show that compared to raw attention, the token attentions from attention rollout and
attention flow have higher correlations with the importance scores obtained from input gradients as | 2005.00928 | Quantifying Attention Flow in Transformers | In the Transformer model, "self-attention" combines information from attended
embeddings into the representation of the focal embedding in the next layer.
Thus, across layers of the Transformer, information originating from different
tokens gets increasingly mixed. This makes attention weights unreliable as
explanations probes. In this paper, we consider the problem of quantifying this
flow of information through self-attention. We propose two methods for
approximating the attention to input tokens given attention weights, attention
rollout and attention flow, as post hoc methods when we use attention weights
as the relative relevance of the input tokens. We show that these methods give
complementary views on the flow of information, and compared to raw attention,
both yield higher correlations with importance scores of input tokens obtained
using an ablation method and input gradients. | http://arxiv.org/pdf/2005.00928 | [
"Samira Abnar",
"Willem Zuidema"
] | [
"cs.LG",
"cs.AI",
"cs.CL"
] | null | null | cs.LG | 20200502 | 20200531 | [] |
2005.00928 | 3 | embedding. We show that compared to raw attention, the token attentions from attention rollout and
attention flow have higher correlations with the importance scores obtained from input gradients as
well as blank-out , an input ablation based attribution method. Furthermore, we visualize the token
attention weights and demonstrate that they are better approximations of how input tokens contribute
to a predicted output, compared to raw attention.
It is noteworthy that the techniques we propose
in this paper, are not toward making hidden embeddings more identifiable, or providing better attention weights for better performance, but a new set
of attention weights that take token identity problem into consideration and can serve as a better
diagnostic tool for visualization and debugging.arXiv:2005.00928v2 [cs.LG] 31 May 2020
(a) Embedding attentions
(b) Attention rollout
(c) Attention flow
Figure 1: Visualisation of attention weights.
Figure 2: Raw Attention maps for the CLS token at
different layers.
2 Setups and Problem Statement
In our analysis, we focus on the verb number prediction task, i.e., predicting singularity or plurality
of a verb of a sentence, when the input is the sentence up to the verb position. We use the subjectverb agreement dataset (Linzen et al., 2016). This | 2005.00928 | Quantifying Attention Flow in Transformers | In the Transformer model, "self-attention" combines information from attended
embeddings into the representation of the focal embedding in the next layer.
Thus, across layers of the Transformer, information originating from different
tokens gets increasingly mixed. This makes attention weights unreliable as
explanations probes. In this paper, we consider the problem of quantifying this
flow of information through self-attention. We propose two methods for
approximating the attention to input tokens given attention weights, attention
rollout and attention flow, as post hoc methods when we use attention weights
as the relative relevance of the input tokens. We show that these methods give
complementary views on the flow of information, and compared to raw attention,
both yield higher correlations with importance scores of input tokens obtained
using an ablation method and input gradients. | http://arxiv.org/pdf/2005.00928 | [
"Samira Abnar",
"Willem Zuidema"
] | [
"cs.LG",
"cs.AI",
"cs.CL"
] | null | null | cs.LG | 20200502 | 20200531 | [] |
2005.00928 | 4 | of a verb of a sentence, when the input is the sentence up to the verb position. We use the subjectverb agreement dataset (Linzen et al., 2016). This
task and dataset are convenient choices, as they offer a clear hypothesis about what part of the input
is essential to get the right solution. For instance,
given “ the key to the cabinets ” as the input, we
know that attending to “key” helps the model predict singular as output while attending to “cabinets”
(anagreement attractor , with the opposite number)
is unhelpful.
We train a Transformer encoder, with GPT2 Transformer blocks as described in (Radford
et al., 2019; Wolf et al., 2019) (without masking).
The model has 6 layers, and 8 heads, with hidden/embedding size of 128. Similar to Bert (Devlin et al., 2019) we add a CLS token and use
its embedding in the final layer as the input to
the classifier. The accuracy of the model on the
subject-verb agreement task is 0:96. To facilitate
replication of our experiments we will make the
implementations of the models we use and algorithms we introduce publicly available at https:
//github.com/samiraabnar/attention_flow . | 2005.00928 | Quantifying Attention Flow in Transformers | In the Transformer model, "self-attention" combines information from attended
embeddings into the representation of the focal embedding in the next layer.
Thus, across layers of the Transformer, information originating from different
tokens gets increasingly mixed. This makes attention weights unreliable as
explanations probes. In this paper, we consider the problem of quantifying this
flow of information through self-attention. We propose two methods for
approximating the attention to input tokens given attention weights, attention
rollout and attention flow, as post hoc methods when we use attention weights
as the relative relevance of the input tokens. We show that these methods give
complementary views on the flow of information, and compared to raw attention,
both yield higher correlations with importance scores of input tokens obtained
using an ablation method and input gradients. | http://arxiv.org/pdf/2005.00928 | [
"Samira Abnar",
"Willem Zuidema"
] | [
"cs.LG",
"cs.AI",
"cs.CL"
] | null | null | cs.LG | 20200502 | 20200531 | [] |
2005.00928 | 5 | implementations of the models we use and algorithms we introduce publicly available at https:
//github.com/samiraabnar/attention_flow .
We start by visualizing raw attention in Figure 1a
(like Vig 2019). The example given here is correctly classified. Crucially, only in the first couple
of layers, there are some distinctions in the attention patterns for different positions, while in higher
layers the attention weights are rather uniform. Fig-ure 2 (left) gives raw attention scores of the CLS
token over input tokens (x-axis) at different layers (y-axis), which similarly lack an interpretable
pattern.These observations reflect the fact that as
we go deeper into the model, the embeddings are
more contextualized and may all carry similar information. This underscores the need to track down
attention weights all the way back to the input layer
and is in line with findings of Serrano and Smith
(2019), who show that attention weights do not
necessarily correspond to the relative importance
of input tokens.
To quantify the usefulness of raw attention
weights, and the two alternatives that we consider
in the next section, besides input gradients, we
employ an input ablation method, blank-out , to estimate an importance score for each input token. | 2005.00928 | Quantifying Attention Flow in Transformers | In the Transformer model, "self-attention" combines information from attended
embeddings into the representation of the focal embedding in the next layer.
Thus, across layers of the Transformer, information originating from different
tokens gets increasingly mixed. This makes attention weights unreliable as
explanations probes. In this paper, we consider the problem of quantifying this
flow of information through self-attention. We propose two methods for
approximating the attention to input tokens given attention weights, attention
rollout and attention flow, as post hoc methods when we use attention weights
as the relative relevance of the input tokens. We show that these methods give
complementary views on the flow of information, and compared to raw attention,
both yield higher correlations with importance scores of input tokens obtained
using an ablation method and input gradients. | http://arxiv.org/pdf/2005.00928 | [
"Samira Abnar",
"Willem Zuidema"
] | [
"cs.LG",
"cs.AI",
"cs.CL"
] | null | null | cs.LG | 20200502 | 20200531 | [] |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.