{ "id": 298081, "title": "5th place solution, maskrcnn, pseudo label and cellpose", "url": "/competitions/sartorius-cell-instance-segmentation/writeups/bestfitting-5th-place-solution-maskrcnn-pseudo-lab", "author_name": "", "post_date": "2022-02-24T03:52:37.870Z", "votes": 84, "comment_count": 15, "views": 0, "content": "

Congrats to all the winners, and thanks to Sartorius team and Kaggle hosted this competition, it’s very challenging and interest.

\n

Methods

\n

My pipeline can be described in the following figure.
\n\"pipeline.png\"
\nStep 1: MaskRCNN models were trained on Detectron2 using LIVECell_dataset_2021 without shsy5y cells.
\nStep 2: Finetune the MaskRCNN models on competition train dataset and LIVECell shsy5y cells.
\nStep 3: Predict the train_semi_supervised dataset and used them as pseudo labels.
\nStep 4: Finetune the models using LIVECellshsy5y cells + competition trainset + train_semi_supervised(pseudo)
\nStep 5: Predict the competition trainset and LIVECellshsy5y cells using the above models.
\nStep 6: Generate the flow-x, flow-y and semantic segmentation base on Step5 results using Cellpose.
\nStep 7: Train Cellpose model on LIVECell_dataset_2021 without shsy5y cells.
\nStep 8: Finetune Cellpose model using Step7 results as additional channels.
\nStep 9: Predict and post-process using Cellpose. The diameter was set to 19 and re-predict.

\n

The MaskRCNN+Cellpose architecture is in the following figure.
\n\"cellpose-arch.png\"
\nResults
\n\"results.png\"
\nThings tried but not worked
\n1.Deep Watershed and it’s variants.
\n2.OmniPose

\n

Things want to try if have more time
\n1.Cellpose as additional heads of MaskRCNN.
\n2.Detection models using all kinds of datasets+UNET Singel Cell segmentation.

\n

Happy New Year!

", "messages": [ { "id": "1634288", "postDate": "12/31/2021 14:26:44", "content": "

Congrats to all the winners, and thanks to Sartorius team and Kaggle hosted this competition, it’s very challenging and interest.

\n

Methods

\n

My pipeline can be described in the following figure.
\n\"pipeline.png\"
\nStep 1: MaskRCNN models were trained on Detectron2 using LIVECell_dataset_2021 without shsy5y cells.
\nStep 2: Finetune the MaskRCNN models on competition train dataset and LIVECell shsy5y cells.
\nStep 3: Predict the train_semi_supervised dataset and used them as pseudo labels.
\nStep 4: Finetune the models using LIVECellshsy5y cells + competition trainset + train_semi_supervised(pseudo)
\nStep 5: Predict the competition trainset and LIVECellshsy5y cells using the above models.
\nStep 6: Generate the flow-x, flow-y and semantic segmentation base on Step5 results using Cellpose.
\nStep 7: Train Cellpose model on LIVECell_dataset_2021 without shsy5y cells.
\nStep 8: Finetune Cellpose model using Step7 results as additional channels.
\nStep 9: Predict and post-process using Cellpose. The diameter was set to 19 and re-predict.

\n

The MaskRCNN+Cellpose architecture is in the following figure.
\n\"cellpose-arch.png\"
\nResults
\n\"results.png\"
\nThings tried but not worked
\n1.Deep Watershed and it’s variants.
\n2.OmniPose

\n

Things want to try if have more time
\n1.Cellpose as additional heads of MaskRCNN.
\n2.Detection models using all kinds of datasets+UNET Singel Cell segmentation.

\n

Happy New Year!

", "rawMarkdown": "Congrats to all the winners, and thanks to Sartorius team and Kaggle hosted this competition, it’s very challenging and interest.\n\n**Methods**\n\nMy pipeline can be described in the following figure.\n![pipeline.png](https://i.imgur.com/pF2p304.png)\nStep 1: MaskRCNN models were trained on Detectron2 using LIVECell_dataset_2021 without shsy5y cells.\nStep 2: Finetune the MaskRCNN models on competition train dataset and LIVECell shsy5y cells.\nStep 3: Predict the train_semi_supervised dataset and used them as pseudo labels.\nStep 4: Finetune the models using LIVECellshsy5y cells + competition trainset + train_semi_supervised(pseudo)\nStep 5: Predict the competition trainset and LIVECellshsy5y cells using the above models.\nStep 6: Generate the flow-x, flow-y and semantic segmentation base on Step5 results using Cellpose.\nStep 7: Train Cellpose model on LIVECell_dataset_2021 without shsy5y cells.\nStep 8: Finetune Cellpose model using Step7 results as additional channels.\nStep 9: Predict and post-process using Cellpose. The diameter was set to 19 and re-predict.\n\nThe MaskRCNN+Cellpose architecture is in the following figure.\n![cellpose-arch.png](https://i.imgur.com/WJnAQOH.png)\n**Results**\n![results.png](https://i.imgur.com/dwA0BNE.png)\n**Things tried but not worked**\n1.Deep Watershed and it’s variants.\n2.OmniPose \n\n**Things want to try if have more time**\n1.Cellpose as additional heads of MaskRCNN.\n2.Detection models using all kinds of datasets+UNET Singel Cell segmentation.\n\n\n**Happy New Year!**", "votes": null }, { "id": "1634332", "postDate": "12/31/2021 15:44:58", "content": "

Happy New Year!

", "rawMarkdown": "Happy New Year!", "votes": null }, { "id": "1634336", "postDate": "12/31/2021 15:53:30", "content": "

Interesting idea with adding maskrcnn output as extra channels! And I see it gave you a good boost in performance. I never thought of that.

\n

Have you not experimented with other diameter setting than constant 19? It was a big factor in my own solution. You already have generated masks you are passing to it so I would have set to the median size of that.

", "rawMarkdown": "Interesting idea with adding maskrcnn output as extra channels! And I see it gave you a good boost in performance. I never thought of that.\n\nHave you not experimented with other diameter setting than constant 19? It was a big factor in my own solution. You already have generated masks you are passing to it so I would have set to the median size of that.", "votes": null }, { "id": "1634354", "postDate": "12/31/2021 16:09:39", "content": "

Nice work @bestfitting. Good to see your solution has swin transformer.

", "rawMarkdown": "Nice work @bestfitting. Good to see your solution has swin transformer.", "votes": null }, { "id": "1634360", "postDate": "12/31/2021 16:17:05", "content": "

Congrats! :)
\nI predict the mask of an image, get the median diameter of the cells, and then re-predict this image using the estimated diameter.

", "rawMarkdown": "Congrats! :)\nI predict the mask of an image, get the median diameter of the cells, and then re-predict this image using the estimated diameter.", "votes": null }, { "id": "1634363", "postDate": "12/31/2021 16:21:11", "content": "

Ah ok, that makes sense.

", "rawMarkdown": "Ah ok, that makes sense.", "votes": null }, { "id": "1634401", "postDate": "12/31/2021 17:04:15", "content": "

I noticed the shsy5y is excluded and then included back into the pipeline. May I ask what was the results/observations that led you to do that?

", "rawMarkdown": "I noticed the shsy5y is excluded and then included back into the pipeline. May I ask what was the results/observations that led you to do that?", "votes": null }, { "id": "1634402", "postDate": "12/31/2021 17:04:29", "content": "

Nice and awesome model ,Congrats @bestfitting

", "rawMarkdown": "Nice and awesome model ,Congrats @bestfitting", "votes": null }, { "id": "1634500", "postDate": "12/31/2021 18:25:14", "content": "

A very articulated and complete work, @bestfitting. Congrats for your excellent result.

", "rawMarkdown": "A very articulated and complete work, @bestfitting. Congrats for your excellent result.", "votes": null }, { "id": "1634550", "postDate": "12/31/2021 19:40:42", "content": "

Amazing engineering work! I'm absolutely impressed by your capacity to merge these multiple frameworks in the same integrated solution. I'm pretty sure that if you had 1 more week you would be top 1 with some more tuning. You had by far the steepest rising curve in the last week.

", "rawMarkdown": "Amazing engineering work! I'm absolutely impressed by your capacity to merge these multiple frameworks in the same integrated solution. I'm pretty sure that if you had 1 more week you would be top 1 with some more tuning. You had by far the steepest rising curve in the last week.", "votes": null }, { "id": "1634889", "postDate": "01/01/2022 09:32:45", "content": "

Thanks for your kind words, but time is fair to everyone, so I quite satisfied with the final result. :)

", "rawMarkdown": "Thanks for your kind words, but time is fair to everyone, so I quite satisfied with the final result. :)", "votes": null }, { "id": "1634895", "postDate": "01/01/2022 09:38:25", "content": "

@jy2tong The logic behind is: I treat shsy5y cells of LIVECell_dataset_2021 as a part of train set data.

", "rawMarkdown": "jy2tong The logic behind is: I treat shsy5y cells of LIVECell_dataset_2021 as a part of train set data.", "votes": null }, { "id": "1635385", "postDate": "01/01/2022 17:58:06", "content": "

Congratulations !great work

", "rawMarkdown": "Congratulations !great work", "votes": null }, { "id": "1635592", "postDate": "01/02/2022 00:56:22", "content": "

Thank you.

", "rawMarkdown": "Thank you.", "votes": null }, { "id": "1649323", "postDate": "01/14/2022 07:27:06", "content": "

How do you generate three channels from the Instance segmentation?

", "rawMarkdown": "How do you generate three channels from the Instance segmentation?", "votes": null }, { "id": "1656519", "postDate": "01/19/2022 11:56:38", "content": "

Eager to open source learning.

", "rawMarkdown": "Eager to open source learning.", "votes": null } ], "comments": [ { "id": 1634332, "author_name": "tiandaye", "author_url": "", "post_date": "12/31/2021 15:44:58", "content": "

Happy New Year!

", "votes": null, "replies": [] }, { "id": 1634336, "author_name": "slawekbiel", "author_url": "", "post_date": "12/31/2021 15:53:30", "content": "

Interesting idea with adding maskrcnn output as extra channels! And I see it gave you a good boost in performance. I never thought of that.

\n

Have you not experimented with other diameter setting than constant 19? It was a big factor in my own solution. You already have generated masks you are passing to it so I would have set to the median size of that.

", "votes": null, "replies": [ { "id": 1634360, "author_name": "bestfitting", "author_url": "", "post_date": "12/31/2021 16:17:05", "content": "

Congrats! :)
\nI predict the mask of an image, get the median diameter of the cells, and then re-predict this image using the estimated diameter.

", "votes": null, "replies": [] }, { "id": 1634363, "author_name": "slawekbiel", "author_url": "", "post_date": "12/31/2021 16:21:11", "content": "

Ah ok, that makes sense.

", "votes": null, "replies": [] } ] }, { "id": 1634354, "author_name": "duykhanh99", "author_url": "", "post_date": "12/31/2021 16:09:39", "content": "

Nice work @bestfitting. Good to see your solution has swin transformer.

", "votes": null, "replies": [] }, { "id": 1634401, "author_name": "jy2tong", "author_url": "", "post_date": "12/31/2021 17:04:15", "content": "

I noticed the shsy5y is excluded and then included back into the pipeline. May I ask what was the results/observations that led you to do that?

", "votes": null, "replies": [ { "id": 1634895, "author_name": "bestfitting", "author_url": "", "post_date": "01/01/2022 09:38:25", "content": "

@jy2tong The logic behind is: I treat shsy5y cells of LIVECell_dataset_2021 as a part of train set data.

", "votes": null, "replies": [] }, { "id": 1635592, "author_name": "jy2tong", "author_url": "", "post_date": "01/02/2022 00:56:22", "content": "

Thank you.

", "votes": null, "replies": [] } ] }, { "id": 1634402, "author_name": "balavashan", "author_url": "", "post_date": "12/31/2021 17:04:29", "content": "

Nice and awesome model ,Congrats @bestfitting

", "votes": null, "replies": [] }, { "id": 1634500, "author_name": "lucamassaron", "author_url": "", "post_date": "12/31/2021 18:25:14", "content": "

A very articulated and complete work, @bestfitting. Congrats for your excellent result.

", "votes": null, "replies": [] }, { "id": 1634550, "author_name": "alexandrecc", "author_url": "", "post_date": "12/31/2021 19:40:42", "content": "

Amazing engineering work! I'm absolutely impressed by your capacity to merge these multiple frameworks in the same integrated solution. I'm pretty sure that if you had 1 more week you would be top 1 with some more tuning. You had by far the steepest rising curve in the last week.

", "votes": null, "replies": [ { "id": 1634889, "author_name": "bestfitting", "author_url": "", "post_date": "01/01/2022 09:32:45", "content": "

Thanks for your kind words, but time is fair to everyone, so I quite satisfied with the final result. :)

", "votes": null, "replies": [] } ] }, { "id": 1635385, "author_name": "aiswaryasivakumar", "author_url": "", "post_date": "01/01/2022 17:58:06", "content": "

Congratulations !great work

", "votes": null, "replies": [] }, { "id": 1649323, "author_name": "zaopolearning", "author_url": "", "post_date": "01/14/2022 07:27:06", "content": "

How do you generate three channels from the Instance segmentation?

", "votes": null, "replies": [] }, { "id": 1656519, "author_name": "zaopolearning", "author_url": "", "post_date": "01/19/2022 11:56:38", "content": "

Eager to open source learning.

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1634288": "Congrats to all the winners, and thanks to Sartorius team and Kaggle hosted this competition, it’s very challenging and interest.\n\n**Methods**\n\nMy pipeline can be described in the following figure.\n![pipeline.png](https://i.imgur.com/pF2p304.png)\nStep 1: MaskRCNN models were trained on Detectron2 using LIVECell_dataset_2021 without shsy5y cells.\nStep 2: Finetune the MaskRCNN models on competition train dataset and LIVECell shsy5y cells.\nStep 3: Predict the train_semi_supervised dataset and used them as pseudo labels.\nStep 4: Finetune the models using LIVECellshsy5y cells + competition trainset + train_semi_supervised(pseudo)\nStep 5: Predict the competition trainset and LIVECellshsy5y cells using the above models.\nStep 6: Generate the flow-x, flow-y and semantic segmentation base on Step5 results using Cellpose.\nStep 7: Train Cellpose model on LIVECell_dataset_2021 without shsy5y cells.\nStep 8: Finetune Cellpose model using Step7 results as additional channels.\nStep 9: Predict and post-process using Cellpose. The diameter was set to 19 and re-predict.\n\nThe MaskRCNN+Cellpose architecture is in the following figure.\n![cellpose-arch.png](https://i.imgur.com/WJnAQOH.png)\n**Results**\n![results.png](https://i.imgur.com/dwA0BNE.png)\n**Things tried but not worked**\n1.Deep Watershed and it’s variants.\n2.OmniPose \n\n**Things want to try if have more time**\n1.Cellpose as additional heads of MaskRCNN.\n2.Detection models using all kinds of datasets+UNET Singel Cell segmentation.\n\n\n**Happy New Year!**", "1634332": "Happy New Year!", "1634336": "Interesting idea with adding maskrcnn output as extra channels! And I see it gave you a good boost in performance. I never thought of that.\n\nHave you not experimented with other diameter setting than constant 19? It was a big factor in my own solution. You already have generated masks you are passing to it so I would have set to the median size of that.", "1634354": "Nice work @bestfitting. Good to see your solution has swin transformer.", "1634360": "Congrats! :)\nI predict the mask of an image, get the median diameter of the cells, and then re-predict this image using the estimated diameter.", "1634363": "Ah ok, that makes sense.", "1634401": "I noticed the shsy5y is excluded and then included back into the pipeline. May I ask what was the results/observations that led you to do that?", "1634402": "Nice and awesome model ,Congrats @bestfitting", "1634500": "A very articulated and complete work, @bestfitting. Congrats for your excellent result.", "1634550": "Amazing engineering work! I'm absolutely impressed by your capacity to merge these multiple frameworks in the same integrated solution. I'm pretty sure that if you had 1 more week you would be top 1 with some more tuning. You had by far the steepest rising curve in the last week.", "1634889": "Thanks for your kind words, but time is fair to everyone, so I quite satisfied with the final result. :)", "1634895": "jy2tong The logic behind is: I treat shsy5y cells of LIVECell_dataset_2021 as a part of train set data.", "1635385": "Congratulations !great work", "1635592": "Thank you.", "1649323": "How do you generate three channels from the Instance segmentation?", "1656519": "Eager to open source learning." }, "source": "meta" }