\nhttps://www.youtube.com/watch?v=7GVuZJjroMo&t=1050s
\n
![\"https://i.ibb.co/nkk7jCP/Screenshot-2022-01-30-141729.png\"](\"https://i.ibb.co/nkk7jCP/Screenshot-2022-01-30-141729.png\")
i am not taking about tracking.
\nbut rather using frame,t-1,t-2 … for object detection (in last frame).
\ne.g. you use 3d conv head, multi-frame attention head, etc
quote from cots dataset paper:
\nhttps://arxiv.org/pdf/2111.14311.pdf
\n\"The dataset naturally exhibits sequence-based annotations
\nas multiple images are taken of the same COTS as the boat
\nmoves past it.\"
Vertigo3 underwater true-flight glider with application to broad scale COTS management in the GBR - Dr Russ Babcock, Dr Brett Kettle
\nhttps://www.youtube.com/watch?v=7GVuZJjroMo&t=1050s
\n
i am not taking about tracking.
\nbut rather using frame,t-1,t-2 … for object detection (in last frame).
\ne.g. you use 3d conv head, multi-frame attention head, etc
quote from cots dataset paper:
\nhttps://arxiv.org/pdf/2111.14311.pdf
\n\"The dataset naturally exhibits sequence-based annotations
\nas multiple images are taken of the same COTS as the boat
\nmoves past it.\"
Vertigo3 underwater true-flight glider with application to broad scale COTS management in the GBR - Dr Russ Babcock, Dr Brett Kettle
\nhttps://www.youtube.com/watch?v=7GVuZJjroMo&t=1050s
\n
Dear Mr frog , I have another question
\nDo you think gopro wide-camera fisheye and perspective distortion have a result in smaller object around sides?
like this
\nwill de-fisheye or anti-perspect preprocess or according data augments make improve on detection performace?
\nI run over past sompetition find this idea
\nhttps://www.kaggle.com/c/pku-autonomous-driving/discussion/127037
i don't think that is required unless you are trying to predict some 3d values, or you are using 3d features in your solution
\n(for our case, we need only to predict 2d bbox. you can add perspective or affine augmentation in your training)
", "rawMarkdown": "i don't think that is required unless you are trying to predict some 3d values, or you are using 3d features in your solution\n\n(for our case, we need only to predict 2d bbox. you can add perspective or affine augmentation in your training)", "votes": null }, { "id": "1670353", "postDate": "01/31/2022 12:56:06", "content": "I will keep searching， may fortune come for you, happy spring festival
", "rawMarkdown": "I will keep searching， may fortune come for you, happy spring festival", "votes": null }, { "id": "1670369", "postDate": "01/31/2022 13:16:21", "content": "A shortcut and cheating way to test augmentation:
\nwe use classifier to model distribution like GAN. train a classifier (or object detector) for 3 class : train pos, validation pos, all neg. given a pos input sample, we can measure if it is closer to the train or validation set.
let T be the original train set. TA be the train set with augmentation applied. using the trained classifier above, check if TA is closer to the validation set (good augmentation), train set or background set (poor augmentation).
@hengck23 your tracking dataset + https://github.com/open-mmlab/mmtracking = magic
", "rawMarkdown": "hengck23 your tracking dataset + https://github.com/open-mmlab/mmtracking = magic", "votes": null }, { "id": "1670738", "postDate": "01/31/2022 20:24:43", "content": "Temporal RoI Align looks the most promising, had little luck getting it set up because I'm trying to focus on TF2.0 right now. But I have plans to revisit
", "rawMarkdown": "Temporal RoI Align looks the most promising, had little luck getting it set up because I'm trying to focus on TF2.0 right now. But I have plans to revisit", "votes": null } ], "comments": [ { "id": 1669237, "author_name": "drzhuzhe", "author_url": "", "post_date": "01/30/2022 14:01:16", "content": "Dear Mr frog , I have another question
\nDo you think gopro wide-camera fisheye and perspective distortion have a result in smaller object around sides?
like this
\nwill de-fisheye or anti-perspect preprocess or according data augments make improve on detection performace?
\nI run over past sompetition find this idea
\nhttps://www.kaggle.com/c/pku-autonomous-driving/discussion/127037
i don't think that is required unless you are trying to predict some 3d values, or you are using 3d features in your solution
\n(for our case, we need only to predict 2d bbox. you can add perspective or affine augmentation in your training)
", "votes": null, "replies": [] }, { "id": 1670353, "author_name": "drzhuzhe", "author_url": "", "post_date": "01/31/2022 12:56:06", "content": "I will keep searching， may fortune come for you, happy spring festival
", "votes": null, "replies": [] }, { "id": 1670369, "author_name": "hengck23", "author_url": "", "post_date": "01/31/2022 13:16:21", "content": "A shortcut and cheating way to test augmentation:
\nwe use classifier to model distribution like GAN. train a classifier (or object detector) for 3 class : train pos, validation pos, all neg. given a pos input sample, we can measure if it is closer to the train or validation set.
let T be the original train set. TA be the train set with augmentation applied. using the trained classifier above, check if TA is closer to the validation set (good augmentation), train set or background set (poor augmentation).
@hengck23 your tracking dataset + https://github.com/open-mmlab/mmtracking = magic
", "votes": null, "replies": [ { "id": 1670738, "author_name": "outwrest", "author_url": "", "post_date": "01/31/2022 20:24:43", "content": "Temporal RoI Align looks the most promising, had little luck getting it set up because I'm trying to focus on TF2.0 right now. But I have plans to revisit
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1668888": "i am not taking about tracking.\n\nbut rather using frame,t-1,t-2 ... for object detection (in last frame).\ne.g. you use 3d conv head, multi-frame attention head, etc\n\nquote from cots dataset paper:\nhttps://arxiv.org/pdf/2111.14311.pdf\n\"The dataset naturally exhibits sequence-based annotations\nas multiple images are taken of the same COTS as the boat\nmoves past it.\"\n\n\nVertigo3 underwater true-flight glider with application to broad scale COTS management in the GBR - Dr Russ Babcock, Dr Brett Kettle\nhttps://www.youtube.com/watch?v=7GVuZJjroMo&t=1050s\n", "1669237": "Dear Mr frog , I have another question\nDo you think gopro wide-camera fisheye and perspective distortion have a result in smaller object around sides?\n\nlike this \n\n![perspect](https://imgur.com/bNABgvS \"\")\n\n![anti perspect](https://imgur.com/Xe396jz \"\")\n\nwill de-fisheye or anti-perspect preprocess or according data augments make improve on detection performace?\n\n\nI run over past sompetition find this idea \nhttps://www.kaggle.com/c/pku-autonomous-driving/discussion/127037", "1670229": "i don't think that is required unless you are trying to predict some 3d values, or you are using 3d features in your solution\n\n(for our case, we need only to predict 2d bbox. you can add perspective or affine augmentation in your training)", "1670353": "I will keep searching， may fortune come for you, happy spring festival", "1670369": "A shortcut and cheating way to test augmentation:\n\n1. we use classifier to model distribution like GAN. train a classifier (or object detector) for 3 class : train pos, validation pos, all neg. given a pos input sample, we can measure if it is closer to the train or validation set.\n\n2. let T be the original train set. TA be the train set with augmentation applied. using the trained classifier above, check if TA is closer to the validation set (good augmentation), train set or background set (poor augmentation).", "1670736": "hengck23 your tracking dataset + https://github.com/open-mmlab/mmtracking = magic", "1670738": "Temporal RoI Align looks the most promising, had little luck getting it set up because I'm trying to focus on TF2.0 right now. But I have plans to revisit" }, "source": "meta" }![\"perspect\"](\"https://imgur.com/bNABgvS\")
![\"anti](\"https://imgur.com/Xe396jz\")