{ "id": 297998, "title": "8th place solution", "url": "/competitions/sartorius-cell-instance-segmentation/writeups/odede-8th-place-solution", "author_name": "", "post_date": "2021-12-31T09:33:39.963Z", "votes": 54, "comment_count": 7, "views": 0, "content": "

My solution points

\n
    \n
  1. cascade mask rcnn with resnext152 backbone
    2. \n
  2. pretrain on LIVECell dataset
    3. \n
  3. pseudo labeling
    4. \n
  4. WBF and WMF(weighted masks fusion)
    5. \n
  5. train and inference on each image class
    6. \n
\n

Pipeline is this

\n

Training strategy :
\nI tried to train the model in all classes, but it did not work, so I trained and inferenced with each class.
\nI pretrained models on LIVECell dataset and finetuned with competition data. Then, I inferenced on train-semi-supervised data to generate pseudo labels. And finally finetuned these models with competition data and pseudo labels.
\nThe ways to pretrain and finetune are similar to these codes. pretrain, finetune

\n

Inference strategy:
\nI ensemble the boxes predicted by the cascade mask rcnn and yolov5x with WBF, and use the boxes to generate masks. I use WMF, which is WBF applied to the mask ensemble, and ensemble the masks of folds.
\nInference code and WMF code is here.Inference code, WMF code
\nWMF code is identical to this except for including WMF.

\n

I have spent hundreds of hours on this competition, and I am very happy with the results!!

", "messages": [ { "id": "1633731", "postDate": "12/31/2021 02:27:01", "content": "

My solution points

\n
    \n
  1. cascade mask rcnn with resnext152 backbone
    2. \n
  2. pretrain on LIVECell dataset
    3. \n
  3. pseudo labeling
    4. \n
  4. WBF and WMF(weighted masks fusion)
    5. \n
  5. train and inference on each image class
    6. \n
\n

Pipeline is this

\n

Training strategy :
\nI tried to train the model in all classes, but it did not work, so I trained and inferenced with each class.
\nI pretrained models on LIVECell dataset and finetuned with competition data. Then, I inferenced on train-semi-supervised data to generate pseudo labels. And finally finetuned these models with competition data and pseudo labels.
\nThe ways to pretrain and finetune are similar to these codes. pretrain, finetune

\n

Inference strategy:
\nI ensemble the boxes predicted by the cascade mask rcnn and yolov5x with WBF, and use the boxes to generate masks. I use WMF, which is WBF applied to the mask ensemble, and ensemble the masks of folds.
\nInference code and WMF code is here.Inference code, WMF code
\nWMF code is identical to this except for including WMF.

\n

I have spent hundreds of hours on this competition, and I am very happy with the results!!

", "rawMarkdown": "My solution points\n1. cascade mask rcnn with resnext152 backbone\n2. pretrain on LIVECell dataset\n3. pseudo labeling\n4. WBF and WMF(weighted masks fusion)\n5. train and inference on each image class\n\nPipeline is [this](https://drive.google.com/file/d/1Bl5AzzTsm9_tjwezlXUvyEprOwmZcWSw/view?usp=sharing)\n\nTraining strategy : \nI tried to train the model in all classes, but it did not work, so I trained and inferenced with each class.\nI pretrained models on LIVECell dataset and finetuned with competition data. Then, I inferenced on train-semi-supervised data to generate pseudo labels. And finally finetuned these models with competition data and pseudo labels.\nThe ways to pretrain and finetune are similar to these codes. [pretrain](https://www.kaggle.com/markunys/sartorius-transfer-learning-train-with-livecell), [finetune](https://www.kaggle.com/markunys/sartorius-transfer-learning-train)\n\nInference strategy:\nI ensemble the boxes predicted by the cascade mask rcnn and yolov5x with WBF, and use the boxes to generate masks. I use WMF, which is WBF applied to the mask ensemble, and ensemble the masks of folds.\nInference code and WMF code is here.[Inference code](https://www.kaggle.com/markunys/8th-place-solution-inference), [WMF code](https://www.kaggle.com/markunys/ensemble-boxes)\nWMF code is identical to [this](https://github.com/ZFTurbo/Weighted-Boxes-Fusion) except for including WMF.\n\nI have spent hundreds of hours on this competition, and I am very happy with the results!!", "votes": null }, { "id": "1633752", "postDate": "12/31/2021 03:07:12", "content": "

Congrats on solo gold @markunys! Nice work and thanks for sharing.

", "rawMarkdown": "Congrats on solo gold @markunys! Nice work and thanks for sharing.", "votes": null }, { "id": "1633784", "postDate": "12/31/2021 04:07:21", "content": "

Congrats…

", "rawMarkdown": "Congrats...", "votes": null }, { "id": "1633843", "postDate": "12/31/2021 05:14:45", "content": "

Thanks for sharing! Seems like many top solutions tried to ensemble multiple box detector, it really make sense now that we think about it, but have really been a blind spot during the competition!

", "rawMarkdown": "Thanks for sharing! Seems like many top solutions tried to ensemble multiple box detector, it really make sense now that we think about it, but have really been a blind spot during the competition!", "votes": null }, { "id": "1634021", "postDate": "12/31/2021 09:18:13", "content": "

Congratulations and thanks for sharing!

", "rawMarkdown": "Congratulations and thanks for sharing!", "votes": null }, { "id": "1634023", "postDate": "12/31/2021 09:19:43", "content": "

Great hard work and splendid results with your solo gold!

", "rawMarkdown": "Great hard work and splendid results with your solo gold!", "votes": null }, { "id": "1634039", "postDate": "12/31/2021 09:53:51", "content": "

Congratulations! Well deserved gold result given the incredible amount of time you have spent on the competition. Thank you a lot for sharing your explanation, code and pipeline!

", "rawMarkdown": "Congratulations! Well deserved gold result given the incredible amount of time you have spent on the competition. Thank you a lot for sharing your explanation, code and pipeline!", "votes": null }, { "id": "1634730", "postDate": "01/01/2022 05:00:22", "content": "

thanks for sharing ur idea

", "rawMarkdown": "thanks for sharing ur idea", "votes": null } ], "comments": [ { "id": 1633752, "author_name": "yousof9", "author_url": "", "post_date": "12/31/2021 03:07:12", "content": "

Congrats on solo gold @markunys! Nice work and thanks for sharing.

", "votes": null, "replies": [] }, { "id": 1633784, "author_name": "raghaw", "author_url": "", "post_date": "12/31/2021 04:07:21", "content": "

Congrats…

", "votes": null, "replies": [] }, { "id": 1633843, "author_name": "woprime", "author_url": "", "post_date": "12/31/2021 05:14:45", "content": "

Thanks for sharing! Seems like many top solutions tried to ensemble multiple box detector, it really make sense now that we think about it, but have really been a blind spot during the competition!

", "votes": null, "replies": [] }, { "id": 1634021, "author_name": "towhidultonmoy", "author_url": "", "post_date": "12/31/2021 09:18:13", "content": "

Congratulations and thanks for sharing!

", "votes": null, "replies": [] }, { "id": 1634023, "author_name": "lucamassaron", "author_url": "", "post_date": "12/31/2021 09:19:43", "content": "

Great hard work and splendid results with your solo gold!

", "votes": null, "replies": [] }, { "id": 1634039, "author_name": "lucamassaron", "author_url": "", "post_date": "12/31/2021 09:53:51", "content": "

Congratulations! Well deserved gold result given the incredible amount of time you have spent on the competition. Thank you a lot for sharing your explanation, code and pipeline!

", "votes": null, "replies": [] }, { "id": 1634730, "author_name": "aiswaryasivakumar", "author_url": "", "post_date": "01/01/2022 05:00:22", "content": "

thanks for sharing ur idea

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1633731": "My solution points\n1. cascade mask rcnn with resnext152 backbone\n2. pretrain on LIVECell dataset\n3. pseudo labeling\n4. WBF and WMF(weighted masks fusion)\n5. train and inference on each image class\n\nPipeline is [this](https://drive.google.com/file/d/1Bl5AzzTsm9_tjwezlXUvyEprOwmZcWSw/view?usp=sharing)\n\nTraining strategy : \nI tried to train the model in all classes, but it did not work, so I trained and inferenced with each class.\nI pretrained models on LIVECell dataset and finetuned with competition data. Then, I inferenced on train-semi-supervised data to generate pseudo labels. And finally finetuned these models with competition data and pseudo labels.\nThe ways to pretrain and finetune are similar to these codes. [pretrain](https://www.kaggle.com/markunys/sartorius-transfer-learning-train-with-livecell), [finetune](https://www.kaggle.com/markunys/sartorius-transfer-learning-train)\n\nInference strategy:\nI ensemble the boxes predicted by the cascade mask rcnn and yolov5x with WBF, and use the boxes to generate masks. I use WMF, which is WBF applied to the mask ensemble, and ensemble the masks of folds.\nInference code and WMF code is here.[Inference code](https://www.kaggle.com/markunys/8th-place-solution-inference), [WMF code](https://www.kaggle.com/markunys/ensemble-boxes)\nWMF code is identical to [this](https://github.com/ZFTurbo/Weighted-Boxes-Fusion) except for including WMF.\n\nI have spent hundreds of hours on this competition, and I am very happy with the results!!", "1633752": "Congrats on solo gold @markunys! Nice work and thanks for sharing.", "1633784": "Congrats...", "1633843": "Thanks for sharing! Seems like many top solutions tried to ensemble multiple box detector, it really make sense now that we think about it, but have really been a blind spot during the competition!", "1634021": "Congratulations and thanks for sharing!", "1634023": "Great hard work and splendid results with your solo gold!", "1634039": "Congratulations! Well deserved gold result given the incredible amount of time you have spent on the competition. Thank you a lot for sharing your explanation, code and pipeline!", "1634730": "thanks for sharing ur idea" }, "source": "meta" }