https://arxiv.org/abs/1809.00778
\n\nHere is our technical report for team PFDet's 2nd place solution.
\n\nP.S. We're hiring: https://www.preferred-networks.jp/en/jobs
", "messages": [ { "id": "381679", "postDate": "09/05/2018 01:36:55", "content": "https://arxiv.org/abs/1809.00778
\n\nHere is our technical report for team PFDet's 2nd place solution.
\n\nP.S. We're hiring: https://www.preferred-networks.jp/en/jobs
", "rawMarkdown": "https://arxiv.org/abs/1809.00778\n\nHere is our technical report for team PFDet's 2nd place solution.\n\nP.S. We're hiring: https://www.preferred-networks.jp/en/jobs", "votes": null }, { "id": "381762", "postDate": "09/05/2018 06:35:36", "content": "Congrats! Now that the competition is over, I bet you are dying to get rid of this computer that takes up all your living room! ;)
", "rawMarkdown": "Congrats! Now that the competition is over, I bet you are dying to get rid of this computer that takes up all your living room! ;)", "votes": null }, { "id": "381773", "postDate": "09/05/2018 07:16:24", "content": "Congratulation for the excellent works! Aftering reading your technical report, I am intrested in the supressing method you use during the test time. Could you share how you use the non-maximum weighted to replace the non-maximum\nsuppression?
", "rawMarkdown": "Congratulation for the excellent works! Aftering reading your technical report, I am intrested in the supressing method you use during the test time. Could you share how you use the non-maximum weighted to replace the non-maximum\nsuppression?", "votes": null }, { "id": "381778", "postDate": "09/05/2018 07:29:01", "content": "We simply call NMW instead of NMS to suppress duplicate detections. You can think of it as an extension of NMS that uses a weighted average for voting for the position of the final bounding box. NMW is not our idea, but it was proposed in the following ICCV 2017 workshop paper: http://openaccess.thecvf.com/content_ICCV_2017_workshops/papers/w14/Zhou_CAD_Scale_Invariant_ICCV_2017_paper.pdf\nPlease see Section 3.4 for details of how it works.
", "rawMarkdown": "We simply call NMW instead of NMS to suppress duplicate detections. You can think of it as an extension of NMS that uses a weighted average for voting for the position of the final bounding box. NMW is not our idea, but it was proposed in the following ICCV 2017 workshop paper: http://openaccess.thecvf.com/content_ICCV_2017_workshops/papers/w14/Zhou_CAD_Scale_Invariant_ICCV_2017_paper.pdf\nPlease see Section 3.4 for details of how it works.", "votes": null }, { "id": "381792", "postDate": "09/05/2018 07:49:49", "content": "Thx, seems that this post processing method can make use of all the bounding box. I use a similar voting method followed by NMS as well. However, I only take those boxes with iou over a specific threshold to vote for the box adjustment. Our experiment show that this post processing gains nearly 1 point higher on the public LB. Could you share about the NMW improvement on the accuracy.\nFYI, the bounding box voting could be found here:\nhttps://arxiv.org/pdf/1505.01749.pdf
\n\nYou can find the details in Section 4
", "rawMarkdown": "Thx, seems that this post processing method can make use of all the bounding box. I use a similar voting method followed by NMS as well. However, I only take those boxes with iou over a specific threshold to vote for the box adjustment. Our experiment show that this post processing gains nearly 1 point higher on the public LB. Could you share about the NMW improvement on the accuracy.\nFYI, the bounding box voting could be found here:\n[https://arxiv.org/pdf/1505.01749.pdf][1]\n\n\n [1]: https://arxiv.org/pdf/1505.01749.pdf\n\nYou can find the details in Section 4", "votes": null }, { "id": "381838", "postDate": "09/05/2018 09:20:13", "content": "Thanks for the interesting paper!\nWith NMW, we got around 0.65 mAP improvement back when we got around 47- 48 mAP on the public LB, but I'm afraid I don't have any more recent ablative results at the moment.
", "rawMarkdown": "Thanks for the interesting paper!\nWith NMW, we got around 0.65 mAP improvement back when we got around 47- 48 mAP on the public LB, but I'm afraid I don't have any more recent ablative results at the moment.", "votes": null }, { "id": "381941", "postDate": "09/05/2018 12:27:23", "content": "Congratulations both on the result and on a very well written and clear paper!
", "rawMarkdown": "Congratulations both on the result and on a very well written and clear paper!", "votes": null }, { "id": "381960", "postDate": "09/05/2018 13:03:10", "content": "And they were complaining about enormous GPU boxes we've been using back then for Carvana and such. Ha-ha, who could compete with 512 GPU cluster? Just kidding, good work and thank you for writing a paper so fast :)
\n\nP.S. Are you kaggling as part of the work responsibilities or was it like 'anybody here has access to unlimited resources and free to use it in free time'? :)
", "rawMarkdown": "And they were complaining about enormous GPU boxes we've been using back then for Carvana and such. Ha-ha, who could compete with 512 GPU cluster? Just kidding, good work and thank you for writing a paper so fast :) \n\nP.S. Are you kaggling as part of the work responsibilities or was it like 'anybody here has access to unlimited resources and free to use it in free time'? :)", "votes": null }, { "id": "382001", "postDate": "09/05/2018 14:08:29", "content": "Well, we just got a brand-new cluster, and thought we'd take it out for a spin. ;)\nhttps://www.preferred-networks.jp/en/news/pr20180328
\n\nP.S. We're hiring: https://www.preferred-networks.jp/en/jobs
", "rawMarkdown": "Well, we just got a brand-new cluster, and thought we'd take it out for a spin. ;)\nhttps://www.preferred-networks.jp/en/news/pr20180328\n\nP.S. We're hiring: https://www.preferred-networks.jp/en/jobs", "votes": null }, { "id": "382030", "postDate": "09/05/2018 15:20:22", "content": "I am wondering if there is a version 2, which would explain how you implement the Co-occurrence Loss. Because I am confused about use it in the RPN or in the head.
", "rawMarkdown": "I am wondering if there is a version 2, which would explain how you implement the Co-occurrence Loss. Because I am confused about use it in the RPN or in the head.", "votes": null }, { "id": "383100", "postDate": "09/07/2018 17:08:43", "content": "Would it be possible to share the single best model (and eventually some expert models) and put an inferance example using Chainer on github?
", "rawMarkdown": "Would it be possible to share the single best model (and eventually some expert models) and put an inferance example using Chainer on github?", "votes": null }, { "id": "383939", "postDate": "09/10/2018 01:52:57", "content": "Thanks for your question. We use co-occurrence loss only for the head network. E.g. if 'person' is annotated, but the network predicts 'human face' within the area of the 'person' ground-truth, we ignore the prediction when calculating the loss.
", "rawMarkdown": "Thanks for your question. We use co-occurrence loss only for the head network. E.g. if 'person' is annotated, but the network predicts 'human face' within the area of the 'person' ground-truth, we ignore the prediction when calculating the loss.", "votes": null }, { "id": "383941", "postDate": "09/10/2018 01:54:16", "content": "Thanks for the request. We are currently discussing this internally whether to make it public later on.
", "rawMarkdown": "Thanks for the request. We are currently discussing this internally whether to make it public later on.", "votes": null }, { "id": "383985", "postDate": "09/10/2018 04:40:43", "content": "Congratulations! I was really curious about your work because it was on the top of the leaderboard for most of the times.\nThank you for sharing your work!
", "rawMarkdown": "Congratulations! I was really curious about your work because it was on the top of the leaderboard for most of the times.\nThank you for sharing your work!", "votes": null }, { "id": "387922", "postDate": "09/15/2018 23:03:41", "content": "so, as they mentioned in the paper... they had 512 GPU s!
\n\nSo, this is the competition among rich people.
\n\nanyway congrats
", "rawMarkdown": "so, as they mentioned in the paper... they had 512 GPU s!\n\nSo, this is the competition among rich people. \n\nanyway congrats", "votes": null }, { "id": "390261", "postDate": "09/19/2018 23:56:52", "content": "Out of curiosity... how did the dispute over the 1st place ended?
", "rawMarkdown": "Out of curiosity... how did the dispute over the 1st place ended?", "votes": null }, { "id": "463068", "postDate": "01/29/2019 10:54:25", "content": "Hi, I was wondering if your team decide to share the model? I have a problem t hand and was thinking to use your model.
", "rawMarkdown": "Hi, I was wondering if your team decide to share the model? I have a problem t hand and was thinking to use your model.", "votes": null }, { "id": "544093", "postDate": "06/05/2019 05:45:24", "content": "great job, the paper was a good read
", "rawMarkdown": "great job, the paper was a good read", "votes": null } ], "comments": [ { "id": 381762, "author_name": "moshel", "author_url": "", "post_date": "09/05/2018 06:35:36", "content": "Congrats! Now that the competition is over, I bet you are dying to get rid of this computer that takes up all your living room! ;)
", "votes": null, "replies": [] }, { "id": 381773, "author_name": "byryyy", "author_url": "", "post_date": "09/05/2018 07:16:24", "content": "Congratulation for the excellent works! Aftering reading your technical report, I am intrested in the supressing method you use during the test time. Could you share how you use the non-maximum weighted to replace the non-maximum\nsuppression?
", "votes": null, "replies": [] }, { "id": 381778, "author_name": "tkerola", "author_url": "", "post_date": "09/05/2018 07:29:01", "content": "We simply call NMW instead of NMS to suppress duplicate detections. You can think of it as an extension of NMS that uses a weighted average for voting for the position of the final bounding box. NMW is not our idea, but it was proposed in the following ICCV 2017 workshop paper: http://openaccess.thecvf.com/content_ICCV_2017_workshops/papers/w14/Zhou_CAD_Scale_Invariant_ICCV_2017_paper.pdf\nPlease see Section 3.4 for details of how it works.
", "votes": null, "replies": [ { "id": 381792, "author_name": "byryyy", "author_url": "", "post_date": "09/05/2018 07:49:49", "content": "Thx, seems that this post processing method can make use of all the bounding box. I use a similar voting method followed by NMS as well. However, I only take those boxes with iou over a specific threshold to vote for the box adjustment. Our experiment show that this post processing gains nearly 1 point higher on the public LB. Could you share about the NMW improvement on the accuracy.\nFYI, the bounding box voting could be found here:\nhttps://arxiv.org/pdf/1505.01749.pdf
\n\nYou can find the details in Section 4
", "votes": null, "replies": [] }, { "id": 381838, "author_name": "tkerola", "author_url": "", "post_date": "09/05/2018 09:20:13", "content": "Thanks for the interesting paper!\nWith NMW, we got around 0.65 mAP improvement back when we got around 47- 48 mAP on the public LB, but I'm afraid I don't have any more recent ablative results at the moment.
", "votes": null, "replies": [] }, { "id": 382030, "author_name": "byryyy", "author_url": "", "post_date": "09/05/2018 15:20:22", "content": "I am wondering if there is a version 2, which would explain how you implement the Co-occurrence Loss. Because I am confused about use it in the RPN or in the head.
", "votes": null, "replies": [] }, { "id": 383939, "author_name": "tkerola", "author_url": "", "post_date": "09/10/2018 01:52:57", "content": "Thanks for your question. We use co-occurrence loss only for the head network. E.g. if 'person' is annotated, but the network predicts 'human face' within the area of the 'person' ground-truth, we ignore the prediction when calculating the loss.
", "votes": null, "replies": [] } ] }, { "id": 381941, "author_name": "timhartill", "author_url": "", "post_date": "09/05/2018 12:27:23", "content": "Congratulations both on the result and on a very well written and clear paper!
", "votes": null, "replies": [] }, { "id": 381960, "author_name": "ceperaang", "author_url": "", "post_date": "09/05/2018 13:03:10", "content": "And they were complaining about enormous GPU boxes we've been using back then for Carvana and such. Ha-ha, who could compete with 512 GPU cluster? Just kidding, good work and thank you for writing a paper so fast :)
\n\nP.S. Are you kaggling as part of the work responsibilities or was it like 'anybody here has access to unlimited resources and free to use it in free time'? :)
", "votes": null, "replies": [ { "id": 382001, "author_name": "tkerola", "author_url": "", "post_date": "09/05/2018 14:08:29", "content": "Well, we just got a brand-new cluster, and thought we'd take it out for a spin. ;)\nhttps://www.preferred-networks.jp/en/news/pr20180328
\n\nP.S. We're hiring: https://www.preferred-networks.jp/en/jobs
", "votes": null, "replies": [] } ] }, { "id": 383100, "author_name": "lvaleriu", "author_url": "", "post_date": "09/07/2018 17:08:43", "content": "Would it be possible to share the single best model (and eventually some expert models) and put an inferance example using Chainer on github?
", "votes": null, "replies": [ { "id": 383941, "author_name": "tkerola", "author_url": "", "post_date": "09/10/2018 01:54:16", "content": "Thanks for the request. We are currently discussing this internally whether to make it public later on.
", "votes": null, "replies": [] }, { "id": 463068, "author_name": "goelrajat", "author_url": "", "post_date": "01/29/2019 10:54:25", "content": "Hi, I was wondering if your team decide to share the model? I have a problem t hand and was thinking to use your model.
", "votes": null, "replies": [] } ] }, { "id": 383985, "author_name": "tonyapplekim", "author_url": "", "post_date": "09/10/2018 04:40:43", "content": "Congratulations! I was really curious about your work because it was on the top of the leaderboard for most of the times.\nThank you for sharing your work!
", "votes": null, "replies": [] }, { "id": 387922, "author_name": "bellman82", "author_url": "", "post_date": "09/15/2018 23:03:41", "content": "so, as they mentioned in the paper... they had 512 GPU s!
\n\nSo, this is the competition among rich people.
\n\nanyway congrats
", "votes": null, "replies": [] }, { "id": 390261, "author_name": "moshel", "author_url": "", "post_date": "09/19/2018 23:56:52", "content": "Out of curiosity... how did the dispute over the 1st place ended?
", "votes": null, "replies": [] }, { "id": 544093, "author_name": "melunis", "author_url": "", "post_date": "06/05/2019 05:45:24", "content": "great job, the paper was a good read
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "381679": "https://arxiv.org/abs/1809.00778\n\nHere is our technical report for team PFDet's 2nd place solution.\n\nP.S. We're hiring: https://www.preferred-networks.jp/en/jobs", "381762": "Congrats! Now that the competition is over, I bet you are dying to get rid of this computer that takes up all your living room! ;)", "381773": "Congratulation for the excellent works! Aftering reading your technical report, I am intrested in the supressing method you use during the test time. Could you share how you use the non-maximum weighted to replace the non-maximum\nsuppression?", "381778": "We simply call NMW instead of NMS to suppress duplicate detections. You can think of it as an extension of NMS that uses a weighted average for voting for the position of the final bounding box. NMW is not our idea, but it was proposed in the following ICCV 2017 workshop paper: http://openaccess.thecvf.com/content_ICCV_2017_workshops/papers/w14/Zhou_CAD_Scale_Invariant_ICCV_2017_paper.pdf\nPlease see Section 3.4 for details of how it works.", "381792": "Thx, seems that this post processing method can make use of all the bounding box. I use a similar voting method followed by NMS as well. However, I only take those boxes with iou over a specific threshold to vote for the box adjustment. Our experiment show that this post processing gains nearly 1 point higher on the public LB. Could you share about the NMW improvement on the accuracy.\nFYI, the bounding box voting could be found here:\n[https://arxiv.org/pdf/1505.01749.pdf][1]\n\n\n [1]: https://arxiv.org/pdf/1505.01749.pdf\n\nYou can find the details in Section 4", "381838": "Thanks for the interesting paper!\nWith NMW, we got around 0.65 mAP improvement back when we got around 47- 48 mAP on the public LB, but I'm afraid I don't have any more recent ablative results at the moment.", "381941": "Congratulations both on the result and on a very well written and clear paper!", "381960": "And they were complaining about enormous GPU boxes we've been using back then for Carvana and such. Ha-ha, who could compete with 512 GPU cluster? Just kidding, good work and thank you for writing a paper so fast :) \n\nP.S. Are you kaggling as part of the work responsibilities or was it like 'anybody here has access to unlimited resources and free to use it in free time'? :)", "382001": "Well, we just got a brand-new cluster, and thought we'd take it out for a spin. ;)\nhttps://www.preferred-networks.jp/en/news/pr20180328\n\nP.S. We're hiring: https://www.preferred-networks.jp/en/jobs", "382030": "I am wondering if there is a version 2, which would explain how you implement the Co-occurrence Loss. Because I am confused about use it in the RPN or in the head.", "383100": "Would it be possible to share the single best model (and eventually some expert models) and put an inferance example using Chainer on github?", "383939": "Thanks for your question. We use co-occurrence loss only for the head network. E.g. if 'person' is annotated, but the network predicts 'human face' within the area of the 'person' ground-truth, we ignore the prediction when calculating the loss.", "383941": "Thanks for the request. We are currently discussing this internally whether to make it public later on.", "383985": "Congratulations! I was really curious about your work because it was on the top of the leaderboard for most of the times.\nThank you for sharing your work!", "387922": "so, as they mentioned in the paper... they had 512 GPU s!\n\nSo, this is the competition among rich people. \n\nanyway congrats", "390261": "Out of curiosity... how did the dispute over the 1st place ended?", "463068": "Hi, I was wondering if your team decide to share the model? I have a problem t hand and was thinking to use your model.", "544093": "great job, the paper was a good read" }, "source": "meta" }