{ "id": 285422, "title": "Train Attention based Residual Unet with watershed algorithm for instance segmentation", "url": "/competitions/sartorius-cell-instance-segmentation/discussion/285422", "author_name": "", "post_date": "2021-11-04T16:57:39.087727100Z", "votes": 16, "comment_count": 8, "views": 0, "content": "

\n\n

\n

I have created a notebook for Attention-based Residual Unet with some EDA on the three cell types,

\n
\n
    \n
  1. In this notebook I have tried to do a thorough analysis of different cell types given in the dataset.
    2. \n
  2. Did some visualizations for different types of augmentations.
    3. \n
  3. And mainly I tried to Implement Attention based Residual Unet. It is a scratch implementation.
    4. \n
  4. Used tf.data for data pipeline. Tried to make this end to end, from EDA to submission.
    \n But there is a way to perform Instance Segmentation, which is using watershed algorithm with Unet. You can check out this repo which can help you implement that.
    5. \n
\n
", "messages": [ { "id": "1571190", "postDate": "11/04/2021 16:57:39", "content": "

\n\n

\n

I have created a notebook for Attention-based Residual Unet with some EDA on the three cell types,

\n
\n
    \n
  1. In this notebook I have tried to do a thorough analysis of different cell types given in the dataset.
    2. \n
  2. Did some visualizations for different types of augmentations.
    3. \n
  3. And mainly I tried to Implement Attention based Residual Unet. It is a scratch implementation.
    4. \n
  4. Used tf.data for data pipeline. Tried to make this end to end, from EDA to submission.
    \n But there is a way to perform Instance Segmentation, which is using watershed algorithm with Unet. You can check out this repo which can help you implement that.
    5. \n
\n
", "rawMarkdown": "

\n\n

\n \n\nI have created a notebook for [**Attention-based Residual Unet**](https://www.kaggle.com/soumya9977/attention-based-residual-unet-eda-on-cell-imgs) with some EDA on the three cell types, \n\n> 1. In this notebook I have tried to do a thorough analysis of different cell types given in the dataset.\n2. Did some visualizations for different types of augmentations.\n3. And mainly I tried to Implement Attention based Residual Unet. It is a scratch implementation.\n4. Used tf.data for data pipeline. Tried to make this end to end, from EDA to submission.\n\nBut there is a way to perform Instance Segmentation, which is using `watershed algorithm` with Unet. You can check out this [repo](https://github.com/THULiusj/Cell-segmentation-microscope-image-Unet-Watershed-AzureMachineLearningService/tree/master/unet%2Bwatershed_research_code) which can help you implement that.", "votes": null }, { "id": "1574663", "postDate": "11/07/2021 19:09:53", "content": "

I also believe that it is possible to reach a good score using semantic segmentation models like U-Net. In fact, at the moment, my submission is based on that idea.

\n

However, my experiments showed that applying a median filter before dividing a mask into separate masks helps. That's due to a noisy output of a U-Net model.

", "rawMarkdown": "I also believe that it is possible to reach a good score using semantic segmentation models like U-Net. In fact, at the moment, my submission is based on that idea.\n\nHowever, my experiments showed that applying a median filter before dividing a mask into separate masks helps. That's due to a noisy output of a U-Net model.", "votes": null }, { "id": "1577484", "postDate": "11/10/2021 06:08:36", "content": "

Hi @danieliusk have you seen this topic

\n\n

I was thinking that using Unet models wont work, and people were giving me the advice to use MRCNN and stuff, but I guess we can really do some improvements on it.

", "rawMarkdown": "Hi @danieliusk have you seen this `topic`\n- [UNet Strikes Back 🔥 | LB: 0.155+](https://www.kaggle.com/c/sartorius-cell-instance-segmentation/discussion/286553)\n\nI was thinking that using Unet models wont work, and people were giving me the advice to use MRCNN and stuff, but I guess we can really do some improvements on it.", "votes": null }, { "id": "1577489", "postDate": "11/10/2021 06:13:44", "content": "

Sorry for responding late, I started doing other new competitions😅,

\n

You have seen my NB Attention based Residual Unet + EDA on cell imgs🧬 right. I am having a bit of trouble doing the submission, can you please see the submission code from there, and mention what I can improve. seems like every cell boundary prediction will come to a new row for each of the three images.

", "rawMarkdown": "Sorry for responding late, I started doing other new competitions😅,\n\nYou have seen my NB [Attention based Residual Unet + EDA on cell imgs🧬](https://www.kaggle.com/soumya9977/attention-based-residual-unet-eda-on-cell-imgs) right. **I am having a bit of trouble doing the submission, can you please see the submission code from there, and mention what I can improve. seems like every cell boundary prediction will come to a new row for each of the three images.**", "votes": null }, { "id": "1578197", "postDate": "11/10/2021 18:28:04", "content": "

As understand (correct me if I'm wrong) you only predict a binary mask and do not perform any splitting. To begin with, I would suggest using opencv connectedComponents, as it will help you divide mask into separate cell masks.

", "rawMarkdown": "As understand (correct me if I'm wrong) you only predict a binary mask and do not perform any splitting. To begin with, I would suggest using opencv connectedComponents, as it will help you divide mask into separate cell masks.", "votes": null }, { "id": "1580033", "postDate": "11/12/2021 12:40:54", "content": "

Yeah, thanks @danieliusk for your advice, I was able to submit. This time submitted the attention base residual Unet, but the score is not much impressive, it is 0.036. Can you give any advice on how to improve? NB: I did not add any augmentation, adding augmentation might help.

\n\n

Sorry, for asking a lot of questions. I just dont have much experience with kaggle competitions ToT

", "rawMarkdown": "Yeah, thanks @danieliusk for your advice, I was able to submit. This time submitted the `attention base residual Unet`, but the score is not much impressive, it is `0.036`. Can you give any advice on how to improve? NB: I did not add any augmentation, adding augmentation might help.\n\n- I am observing that most of the Unet that are doing well are using pre-trained models. Is there any pretrained version for `attention-based Unet`? \n\n- And another thing if I train my pre-trained Unet model on different backbones then `how can I perform ensemble for all the different backbone-based Unets?`\n\nSorry, for asking a lot of questions. I just dont have much experience with kaggle competitions `ToT`", "votes": null }, { "id": "1580095", "postDate": "11/12/2021 13:20:40", "content": "

I'm glad you were able to submit :)

\n

About improving the model and score:

\n\n

About attention-based U-Net:

\n\n

About ensemble:

\n\n

Hope that helps.

", "rawMarkdown": "I'm glad you were able to submit :) \n\nAbout improving the model and score:\n- I would suggest using full size images for training and using geometric augmentations. I was able to improve my score by using test time augmentations. I'm also thinking about trying different approach (for example segmenting borders, cell centers, etc.) - maybe that could help.\n\nAbout attention-based U-Net:\n- I'm not sure as I'm not using attention based model. My implementation is written from scratch.\n\nAbout ensemble:\n- it's quite easy to ensemble multiple U-Nets if you perform semantic segmentation. One of the simplest ways is to calculate average of outputs you get by inferring same sample through different U-Nets (before splitting mask into different instances).\n\nHope that helps.", "votes": null }, { "id": "1582127", "postDate": "11/14/2021 15:06:40", "content": "

I don't think UNET based models are appropriate for the instance segmentation data in this competition. Most of the instances are extremely close to each other, overlapping at their borders.

", "rawMarkdown": "I don't think UNET based models are appropriate for the instance segmentation data in this competition. Most of the instances are extremely close to each other, overlapping at their borders.", "votes": null }, { "id": "1582389", "postDate": "11/14/2021 20:50:36", "content": "

yeah, @tolgadincer , seems like this is[overlapping issue] a major problem of the solution, this might affect the predictions, I dont know this problem is the same for MRCNN and Detectron too or not. The previous submission of mine was a simple attention-based Residual Unet without any augmentations. But now as @danieliusk suggested I tried TTA with some geometric augmentations at the training time, it seems like that this time there are some overlaps in the predictions, and I'm getting submission errors for this. I am going to try this to remove the overlaps No Overlap Issue: Quick Fix. Lets see where I get.

\n", "rawMarkdown": "yeah, @tolgadincer , seems like this is[overlapping issue] a major problem of the solution, this might affect the predictions, I dont know this problem is the same for MRCNN and Detectron too or not. The previous submission of mine was a simple attention-based Residual Unet without any augmentations. But now as @danieliusk suggested I tried TTA with some geometric augmentations at the training time, it seems like that this time there are some overlaps in the predictions, and I'm getting **submission errors** for this. I am going to try this to remove the overlaps [No Overlap Issue: Quick Fix](https://www.kaggle.com/c/sartorius-cell-instance-segmentation/discussion/279995). Lets see where I get. \n\n- Idea from this `Topic` is not bad either, [UNet Strikes Back 🔥 | LB: 0.155+](https://www.kaggle.com/c/sartorius-cell-instance-segmentation/discussion/286553). Which is to perform ensemble with Unet and MRCNN.\n\n- There are straight Nbs that use straight MRCNN or detectron but still get 0.2+ LB score. Compared to that Unet performance is not that bad, although, I believe that @danieliusk has achieved some good results using Unet only[0.18+ LB], and he is also thinking of improvements.\n\n- I wonder what people are using to get that 0.3+ LB score.", "votes": null } ], "comments": [ { "id": 1574663, "author_name": "danieliusk", "author_url": "", "post_date": "11/07/2021 19:09:53", "content": "

I also believe that it is possible to reach a good score using semantic segmentation models like U-Net. In fact, at the moment, my submission is based on that idea.

\n

However, my experiments showed that applying a median filter before dividing a mask into separate masks helps. That's due to a noisy output of a U-Net model.

", "votes": null, "replies": [ { "id": 1577484, "author_name": "soumya9977", "author_url": "", "post_date": "11/10/2021 06:08:36", "content": "

Hi @danieliusk have you seen this topic

\n\n

I was thinking that using Unet models wont work, and people were giving me the advice to use MRCNN and stuff, but I guess we can really do some improvements on it.

", "votes": null, "replies": [] }, { "id": 1577489, "author_name": "soumya9977", "author_url": "", "post_date": "11/10/2021 06:13:44", "content": "

Sorry for responding late, I started doing other new competitions😅,

\n

You have seen my NB Attention based Residual Unet + EDA on cell imgs🧬 right. I am having a bit of trouble doing the submission, can you please see the submission code from there, and mention what I can improve. seems like every cell boundary prediction will come to a new row for each of the three images.

", "votes": null, "replies": [] }, { "id": 1578197, "author_name": "danieliusk", "author_url": "", "post_date": "11/10/2021 18:28:04", "content": "

As understand (correct me if I'm wrong) you only predict a binary mask and do not perform any splitting. To begin with, I would suggest using opencv connectedComponents, as it will help you divide mask into separate cell masks.

", "votes": null, "replies": [] }, { "id": 1580033, "author_name": "soumya9977", "author_url": "", "post_date": "11/12/2021 12:40:54", "content": "

Yeah, thanks @danieliusk for your advice, I was able to submit. This time submitted the attention base residual Unet, but the score is not much impressive, it is 0.036. Can you give any advice on how to improve? NB: I did not add any augmentation, adding augmentation might help.

\n\n

Sorry, for asking a lot of questions. I just dont have much experience with kaggle competitions ToT

", "votes": null, "replies": [] }, { "id": 1580095, "author_name": "danieliusk", "author_url": "", "post_date": "11/12/2021 13:20:40", "content": "

I'm glad you were able to submit :)

\n

About improving the model and score:

\n\n

About attention-based U-Net:

\n\n

About ensemble:

\n\n

Hope that helps.

", "votes": null, "replies": [] }, { "id": 1582127, "author_name": "tolgadincer", "author_url": "", "post_date": "11/14/2021 15:06:40", "content": "

I don't think UNET based models are appropriate for the instance segmentation data in this competition. Most of the instances are extremely close to each other, overlapping at their borders.

", "votes": null, "replies": [] }, { "id": 1582389, "author_name": "soumya9977", "author_url": "", "post_date": "11/14/2021 20:50:36", "content": "

yeah, @tolgadincer , seems like this is[overlapping issue] a major problem of the solution, this might affect the predictions, I dont know this problem is the same for MRCNN and Detectron too or not. The previous submission of mine was a simple attention-based Residual Unet without any augmentations. But now as @danieliusk suggested I tried TTA with some geometric augmentations at the training time, it seems like that this time there are some overlaps in the predictions, and I'm getting submission errors for this. I am going to try this to remove the overlaps No Overlap Issue: Quick Fix. Lets see where I get.

\n", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1571190": "

\n\n

\n \n\nI have created a notebook for [**Attention-based Residual Unet**](https://www.kaggle.com/soumya9977/attention-based-residual-unet-eda-on-cell-imgs) with some EDA on the three cell types, \n\n> 1. In this notebook I have tried to do a thorough analysis of different cell types given in the dataset.\n2. Did some visualizations for different types of augmentations.\n3. And mainly I tried to Implement Attention based Residual Unet. It is a scratch implementation.\n4. Used tf.data for data pipeline. Tried to make this end to end, from EDA to submission.\n\nBut there is a way to perform Instance Segmentation, which is using `watershed algorithm` with Unet. You can check out this [repo](https://github.com/THULiusj/Cell-segmentation-microscope-image-Unet-Watershed-AzureMachineLearningService/tree/master/unet%2Bwatershed_research_code) which can help you implement that.", "1574663": "I also believe that it is possible to reach a good score using semantic segmentation models like U-Net. In fact, at the moment, my submission is based on that idea.\n\nHowever, my experiments showed that applying a median filter before dividing a mask into separate masks helps. That's due to a noisy output of a U-Net model.", "1577484": "Hi @danieliusk have you seen this `topic`\n- [UNet Strikes Back 🔥 | LB: 0.155+](https://www.kaggle.com/c/sartorius-cell-instance-segmentation/discussion/286553)\n\nI was thinking that using Unet models wont work, and people were giving me the advice to use MRCNN and stuff, but I guess we can really do some improvements on it.", "1577489": "Sorry for responding late, I started doing other new competitions😅,\n\nYou have seen my NB [Attention based Residual Unet + EDA on cell imgs🧬](https://www.kaggle.com/soumya9977/attention-based-residual-unet-eda-on-cell-imgs) right. **I am having a bit of trouble doing the submission, can you please see the submission code from there, and mention what I can improve. seems like every cell boundary prediction will come to a new row for each of the three images.**", "1578197": "As understand (correct me if I'm wrong) you only predict a binary mask and do not perform any splitting. To begin with, I would suggest using opencv connectedComponents, as it will help you divide mask into separate cell masks.", "1580033": "Yeah, thanks @danieliusk for your advice, I was able to submit. This time submitted the `attention base residual Unet`, but the score is not much impressive, it is `0.036`. Can you give any advice on how to improve? NB: I did not add any augmentation, adding augmentation might help.\n\n- I am observing that most of the Unet that are doing well are using pre-trained models. Is there any pretrained version for `attention-based Unet`? \n\n- And another thing if I train my pre-trained Unet model on different backbones then `how can I perform ensemble for all the different backbone-based Unets?`\n\nSorry, for asking a lot of questions. I just dont have much experience with kaggle competitions `ToT`", "1580095": "I'm glad you were able to submit :) \n\nAbout improving the model and score:\n- I would suggest using full size images for training and using geometric augmentations. I was able to improve my score by using test time augmentations. I'm also thinking about trying different approach (for example segmenting borders, cell centers, etc.) - maybe that could help.\n\nAbout attention-based U-Net:\n- I'm not sure as I'm not using attention based model. My implementation is written from scratch.\n\nAbout ensemble:\n- it's quite easy to ensemble multiple U-Nets if you perform semantic segmentation. One of the simplest ways is to calculate average of outputs you get by inferring same sample through different U-Nets (before splitting mask into different instances).\n\nHope that helps.", "1582127": "I don't think UNET based models are appropriate for the instance segmentation data in this competition. Most of the instances are extremely close to each other, overlapping at their borders.", "1582389": "yeah, @tolgadincer , seems like this is[overlapping issue] a major problem of the solution, this might affect the predictions, I dont know this problem is the same for MRCNN and Detectron too or not. The previous submission of mine was a simple attention-based Residual Unet without any augmentations. But now as @danieliusk suggested I tried TTA with some geometric augmentations at the training time, it seems like that this time there are some overlaps in the predictions, and I'm getting **submission errors** for this. I am going to try this to remove the overlaps [No Overlap Issue: Quick Fix](https://www.kaggle.com/c/sartorius-cell-instance-segmentation/discussion/279995). Lets see where I get. \n\n- Idea from this `Topic` is not bad either, [UNet Strikes Back 🔥 | LB: 0.155+](https://www.kaggle.com/c/sartorius-cell-instance-segmentation/discussion/286553). Which is to perform ensemble with Unet and MRCNN.\n\n- There are straight Nbs that use straight MRCNN or detectron but still get 0.2+ LB score. Compared to that Unet performance is not that bad, although, I believe that @danieliusk has achieved some good results using Unet only[0.18+ LB], and he is also thinking of improvements.\n\n- I wonder what people are using to get that 0.3+ LB score." }, "source": "meta" }