{ "id": 214835, "title": "Some resources about weakly-supervised segmentation", "url": "/competitions/hpa-single-cell-image-classification/discussion/214835", "author_name": "xhlulu", "post_date": "2021-01-27T18:59:22.098000", "votes": 60, "comment_count": 4, "views": 0, "content": "
\n

I'm sharing some thoughts based on what I understood from the problem description. If I'm wrong, please correct me and I'll edit this post!

\n
\n

This competition will be particularly interesting and potentially challenging because:

\n
    \n
  1. each training label is assigned to a whole image (weak classification)
    2. \n
  2. evaluation requires predicting the labels+mask for each individual cells in an image (instance segmentation)
    3. \n
\n

Since we don't have access to the masks and individual labels in the training data, this makes the competition more challenging (and thus a more interesting research problem). Below are some resources that might help you get started with weakly-supervised segmentation:

\n\n
\n

Did I miss anything? Please reply below and I'll add it in!

\n
", "messages": [ { "id": 1173267, "postDate": "2021-01-27T18:59:22.097Z", "content": "
\n

I'm sharing some thoughts based on what I understood from the problem description. If I'm wrong, please correct me and I'll edit this post!

\n
\n

This competition will be particularly interesting and potentially challenging because:

\n
    \n
  1. each training label is assigned to a whole image (weak classification)
    2. \n
  2. evaluation requires predicting the labels+mask for each individual cells in an image (instance segmentation)
    3. \n
\n

Since we don't have access to the masks and individual labels in the training data, this makes the competition more challenging (and thus a more interesting research problem). Below are some resources that might help you get started with weakly-supervised segmentation:

\n\n
\n

Did I miss anything? Please reply below and I'll add it in!

\n
", "rawMarkdown": "> I'm sharing some thoughts based on what I *understood* from the problem description. If I'm wrong, please correct me and I'll edit this post!\n\nThis competition will be particularly interesting and potentially challenging because:\n1. each training label is assigned to a whole image (**weak classification**)\n2. evaluation requires predicting the labels+mask for each individual cells in an image (**instance segmentation**)\n\nSince we don't have access to the masks and individual labels in the training data, this makes the competition more challenging (and thus a more interesting research problem). Below are some resources that might help you get started with **weakly-supervised segmentation**:\n\n* If you are not familiar about instance segmentation, check out [those slides](https://www.cs.princeton.edu/courses/archive/spring18/cos598B/public/outline/Instance%20Segmentation.pdf) from a 2018 princeton course.\n* This [section of Papers with Code](https://paperswithcode.com/task/weakly-supervised-semantic-segmentation) gives you a list of popular research papers and source code about weakly-supervised **semantic** segmentation. It's slightly different from **instance** segmentation, and the differences are described in the Princeton slides. \n* Papers with Code also have [a page](https://paperswithcode.com/task/weakly-supervised-instance-segmentation) on **WS instance segmentation** but there's only 4 papers.\n* Want to something more hands-on to learn basic (non-weakly supervised) instance segmentation? Check out this [Keras tutorial](https://www.tensorflow.org/tutorials/images/segmentation) using MobileNet v2 and this [torchvision post](https://pytorch.org/tutorials/intermediate/torchvision_tutorial.html) using Mask R-CNN.\n* Interested in exploring new ideas from broader domain of weakly supervised learning (WSL)? Check out this [ECCV tutorial](https://hbilen.github.io/wsl-eccv20.github.io/) on the subject, and this IEEE paper called *[Weakly Supervised Learning of Single-Cell Feature Embeddings ](https://pubmed.ncbi.nlm.nih.gov/30918435/)*\n\n> Did I miss anything? Please reply below and I'll add it in!", "votes": 59 }, { "id": 1173345, "postDate": "2021-01-27T20:08:21.843Z", "content": "

Nice resource! I just want to point out that the cells can be single-labeled or multi-labeled as well. The single cell labels will often not fully match the image-level labels, plus the addition of Negative class. You can see examples here https://www.kaggle.com/lnhtrang/single-cell-patterns#SCV-and-multi-localization-examples

", "rawMarkdown": "Nice resource! I just want to point out that the cells can be single-labeled or multi-labeled as well. The single cell labels will often not fully match the image-level labels, plus the addition of Negative class. You can see examples here https://www.kaggle.com/lnhtrang/single-cell-patterns#SCV-and-multi-localization-examples ", "votes": 8, "replies": [ { "id": 1176327, "postDate": "2021-01-29T15:24:51.333Z", "content": "

Thanks for sharing this, very helpful!

", "rawMarkdown": "Thanks for sharing this, very helpful!" } ] }, { "id": 1214067, "postDate": "2021-02-22T15:16:10.207Z", "content": "

Hey great resource. Upvoted.

\n

Quick question @xhlulu, the link you provided for the TF/Keras tutorial is just semantic segmentation with a boundary class, object class, and background class… Is it the wrong link or was it just a mistake in thinking it was instance segmentation?

\n

Thanks again!

", "rawMarkdown": "Hey great resource. Upvoted.\n\nQuick question @xhlulu, the link you provided for the TF/Keras tutorial is just semantic segmentation with a boundary class, object class, and background class... Is it the wrong link or was it just a mistake in thinking it was instance segmentation? \n\nThanks again!" }, { "id": 1173664, "postDate": "2021-01-28T03:30:20.480Z", "content": "

Thanks for sharing these resources…

", "rawMarkdown": "Thanks for sharing these resources..." } ], "comments": [ { "id": 1173345, "author_name": "Trang Le", "author_url": "", "post_date": "2021-01-27T20:08:21.843000", "content": "

Nice resource! I just want to point out that the cells can be single-labeled or multi-labeled as well. The single cell labels will often not fully match the image-level labels, plus the addition of Negative class. You can see examples here https://www.kaggle.com/lnhtrang/single-cell-patterns#SCV-and-multi-localization-examples

", "votes": 8, "replies": [ { "id": 1176327, "author_name": "Old Monk", "author_url": "", "post_date": "2021-01-29T15:24:51.333000", "content": "

Thanks for sharing this, very helpful!

", "votes": 0, "replies": [] } ] }, { "id": 1214067, "author_name": "Darien Schettler", "author_url": "", "post_date": "2021-02-22T15:16:10.207000", "content": "

Hey great resource. Upvoted.

\n

Quick question @xhlulu, the link you provided for the TF/Keras tutorial is just semantic segmentation with a boundary class, object class, and background class… Is it the wrong link or was it just a mistake in thinking it was instance segmentation?

\n

Thanks again!

", "votes": 0, "replies": [] }, { "id": 1173664, "author_name": "Gryffindor", "author_url": "", "post_date": "2021-01-28T03:30:20.480000", "content": "

Thanks for sharing these resources…

", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "1173267": "> I'm sharing some thoughts based on what I *understood* from the problem description. If I'm wrong, please correct me and I'll edit this post!\n\nThis competition will be particularly interesting and potentially challenging because:\n1. each training label is assigned to a whole image (**weak classification**)\n2. evaluation requires predicting the labels+mask for each individual cells in an image (**instance segmentation**)\n\nSince we don't have access to the masks and individual labels in the training data, this makes the competition more challenging (and thus a more interesting research problem). Below are some resources that might help you get started with **weakly-supervised segmentation**:\n\n* If you are not familiar about instance segmentation, check out [those slides](https://www.cs.princeton.edu/courses/archive/spring18/cos598B/public/outline/Instance%20Segmentation.pdf) from a 2018 princeton course.\n* This [section of Papers with Code](https://paperswithcode.com/task/weakly-supervised-semantic-segmentation) gives you a list of popular research papers and source code about weakly-supervised **semantic** segmentation. It's slightly different from **instance** segmentation, and the differences are described in the Princeton slides. \n* Papers with Code also have [a page](https://paperswithcode.com/task/weakly-supervised-instance-segmentation) on **WS instance segmentation** but there's only 4 papers.\n* Want to something more hands-on to learn basic (non-weakly supervised) instance segmentation? Check out this [Keras tutorial](https://www.tensorflow.org/tutorials/images/segmentation) using MobileNet v2 and this [torchvision post](https://pytorch.org/tutorials/intermediate/torchvision_tutorial.html) using Mask R-CNN.\n* Interested in exploring new ideas from broader domain of weakly supervised learning (WSL)? Check out this [ECCV tutorial](https://hbilen.github.io/wsl-eccv20.github.io/) on the subject, and this IEEE paper called *[Weakly Supervised Learning of Single-Cell Feature Embeddings ](https://pubmed.ncbi.nlm.nih.gov/30918435/)*\n\n> Did I miss anything? Please reply below and I'll add it in!", "1173345": "Nice resource! I just want to point out that the cells can be single-labeled or multi-labeled as well. The single cell labels will often not fully match the image-level labels, plus the addition of Negative class. You can see examples here https://www.kaggle.com/lnhtrang/single-cell-patterns#SCV-and-multi-localization-examples ", "1214067": "Hey great resource. Upvoted.\n\nQuick question @xhlulu, the link you provided for the TF/Keras tutorial is just semantic segmentation with a boundary class, object class, and background class... Is it the wrong link or was it just a mistake in thinking it was instance segmentation? \n\nThanks again!", "1173664": "Thanks for sharing these resources..." } }