{ "id": 307891, "title": "Any Top Private Subs with Cascade RCNN", "url": "/competitions/tensorflow-great-barrier-reef/discussion/307891", "author_name": "", "post_date": "2022-02-16T05:06:39.341594100Z", "votes": 4, "comment_count": 4, "views": 0, "content": "

I know that @atom1231 was able to get a very high public LB with Cascade RCNN, I wanted to know if anyone was able to get a decently high private with it (since I believe @atom1231 unfortunately dropped out of medal zone).

\n

My other question was if there were any other models (besides @outrunner 's centernet) that was able to outperform YOLOv5 in this competition.

", "messages": [ { "id": "1692524", "postDate": "02/16/2022 05:06:39", "content": "

I know that @atom1231 was able to get a very high public LB with Cascade RCNN, I wanted to know if anyone was able to get a decently high private with it (since I believe @atom1231 unfortunately dropped out of medal zone).

\n

My other question was if there were any other models (besides @outrunner 's centernet) that was able to outperform YOLOv5 in this competition.

", "rawMarkdown": "I know that @atom1231 was able to get a very high public LB with Cascade RCNN, I wanted to know if anyone was able to get a decently high private with it (since I believe @atom1231 unfortunately dropped out of medal zone).\n\nMy other question was if there were any other models (besides @outrunner 's centernet) that was able to outperform YOLOv5 in this competition.", "votes": null }, { "id": "1692558", "postDate": "02/16/2022 05:46:41", "content": "

@kennyxie I was actually working on summarising top solutions and learned the 7th Pos team had a SINGLE MODEL(!!) that was a custom Cascade RCNN. Writeup Link

", "rawMarkdown": "kennyxie I was actually working on summarising [top solutions](https://www.kaggle.com/c/tensorflow-great-barrier-reef/discussion/307831) and learned the 7th Pos team had a SINGLE MODEL(!!) that was a custom Cascade RCNN. [Writeup Link](https://www.kaggle.com/c/tensorflow-great-barrier-reef/discussion/307786)", "votes": null }, { "id": "1692610", "postDate": "02/16/2022 06:23:25", "content": "

My Cascade RCNN R50 (default model) single model got PB 0.69x and it is not well finetuned.
\nI think many models/framework might achieve the top.

", "rawMarkdown": "My Cascade RCNN R50 (default model) single model got PB 0.69x and it is not well finetuned.\nI think many models/framework might achieve the top.", "votes": null }, { "id": "1692622", "postDate": "02/16/2022 06:38:33", "content": "

Will you be writing about your approach?

\n

Sorry if you have and I missed it

", "rawMarkdown": "Will you be writing about your approach?\n\nSorry if you have and I missed it", "votes": null }, { "id": "1692635", "postDate": "02/16/2022 06:48:24", "content": "

train 1280 / inference 3200
\nAugmentation: CLAHE/perspective trasformation/Equalize/motion blur/ColorJitter
\n random crop ratio (0.2~1), then resize back to 1280
\n multiscale 1280-1600
\n (so the valid inference resolution should be 1280+ ~ 6400+)

", "rawMarkdown": "train 1280 / inference 3200 \n Augmentation: CLAHE/perspective trasformation/Equalize/motion blur/ColorJitter\n random crop ratio (0.2~1), then resize back to 1280 \n multiscale 1280-1600\n (so the valid inference resolution should be 1280+ ~ 6400+)", "votes": null } ], "comments": [ { "id": 1692558, "author_name": "init27", "author_url": "", "post_date": "02/16/2022 05:46:41", "content": "

@kennyxie I was actually working on summarising top solutions and learned the 7th Pos team had a SINGLE MODEL(!!) that was a custom Cascade RCNN. Writeup Link

", "votes": null, "replies": [] }, { "id": 1692610, "author_name": "atom1231", "author_url": "", "post_date": "02/16/2022 06:23:25", "content": "

My Cascade RCNN R50 (default model) single model got PB 0.69x and it is not well finetuned.
\nI think many models/framework might achieve the top.

", "votes": null, "replies": [ { "id": 1692622, "author_name": "init27", "author_url": "", "post_date": "02/16/2022 06:38:33", "content": "

Will you be writing about your approach?

\n

Sorry if you have and I missed it

", "votes": null, "replies": [] }, { "id": 1692635, "author_name": "atom1231", "author_url": "", "post_date": "02/16/2022 06:48:24", "content": "

train 1280 / inference 3200
\nAugmentation: CLAHE/perspective trasformation/Equalize/motion blur/ColorJitter
\n random crop ratio (0.2~1), then resize back to 1280
\n multiscale 1280-1600
\n (so the valid inference resolution should be 1280+ ~ 6400+)

", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1692524": "I know that @atom1231 was able to get a very high public LB with Cascade RCNN, I wanted to know if anyone was able to get a decently high private with it (since I believe @atom1231 unfortunately dropped out of medal zone).\n\nMy other question was if there were any other models (besides @outrunner 's centernet) that was able to outperform YOLOv5 in this competition.", "1692558": "kennyxie I was actually working on summarising [top solutions](https://www.kaggle.com/c/tensorflow-great-barrier-reef/discussion/307831) and learned the 7th Pos team had a SINGLE MODEL(!!) that was a custom Cascade RCNN. [Writeup Link](https://www.kaggle.com/c/tensorflow-great-barrier-reef/discussion/307786)", "1692610": "My Cascade RCNN R50 (default model) single model got PB 0.69x and it is not well finetuned.\nI think many models/framework might achieve the top.", "1692622": "Will you be writing about your approach?\n\nSorry if you have and I missed it", "1692635": "train 1280 / inference 3200 \n Augmentation: CLAHE/perspective trasformation/Equalize/motion blur/ColorJitter\n random crop ratio (0.2~1), then resize back to 1280 \n multiscale 1280-1600\n (so the valid inference resolution should be 1280+ ~ 6400+)" }, "source": "meta" }