I have published benchmark notebook, which are using mmdetection instance segmentation model.
\n\nI don't think this model is competitive as a single model, but it may work for ensemble.
\nModel\n mask rcnn\nExtract cells\n I segmented cell masks using CellSegmentator.\nCell Labeling\n labels for each image was assigned to all segmented cells in that image.\nInput colors\n only used RGB, Y is ignored here.\nBackbone\n resnest101\nImage size\n multiscale training 1024-1333\n 1280 for inference\nAugmentation & TTA\n Horizontal FLIP\nScore\n 0.288 on the current macro F1 LB, using 10K (half) training images.\n (I'll train with full data later)\n\n
They are both simple notebooks, so they may be useful for beginners of Instance Segmentation.
\nBy changing the mmdetection configuration file, you can easily experiment with different models.
Have fan!
", "messages": [ { "id": 1182313, "postDate": "2021-02-02T12:08:51.637Z", "content": "I have published benchmark notebook, which are using mmdetection instance segmentation model.
\n\nI don't think this model is competitive as a single model, but it may work for ensemble.
\nModel\n mask rcnn\nExtract cells\n I segmented cell masks using CellSegmentator.\nCell Labeling\n labels for each image was assigned to all segmented cells in that image.\nInput colors\n only used RGB, Y is ignored here.\nBackbone\n resnest101\nImage size\n multiscale training 1024-1333\n 1280 for inference\nAugmentation & TTA\n Horizontal FLIP\nScore\n 0.288 on the current macro F1 LB, using 10K (half) training images.\n (I'll train with full data later)\n\n
They are both simple notebooks, so they may be useful for beginners of Instance Segmentation.
\nBy changing the mmdetection configuration file, you can easily experiment with different models.
Have fan!
", "rawMarkdown": "I have published benchmark notebook, which are using mmdetection instance segmentation model.\n\n* [training](https://www.kaggle.com/its7171/mmdetection-for-segmentation-training)\n* [infernce](https://www.kaggle.com/its7171/mmdetection-for-segmentation-inference)\n\nI don't think this model is competitive as a single model, but it may work for ensemble.\n\n\nModel\n mask rcnn\nExtract cells\n I segmented cell masks using CellSegmentator.\nCell Labeling\n labels for each image was assigned to all segmented cells in that image.\nInput colors\n only used RGB, Y is ignored here.\nBackbone\n resnest101\nImage size\n multiscale training 1024-1333\n 1280 for inference\nAugmentation & TTA\n Horizontal FLIP\nScore\n 0.288 on the current macro F1 LB, using 10K (half) training images.\n (I'll train with full data later)\n\n\nThey are both simple notebooks, so they may be useful for beginners of Instance Segmentation.\nBy changing the mmdetection configuration file, you can easily experiment with different models.\n\nHave fan!", "votes": 36 }, { "id": 1183254, "postDate": "2021-02-02T21:09:07.167Z", "content": "
Hi! Could u clarify on cell labeling please? Do u take into account only the images with 1 label per image? What about the images with multiple labels? Thanks!
", "rawMarkdown": "Hi! Could u clarify on cell labeling please? Do u take into account only the images with 1 label per image? What about the images with multiple labels? Thanks!", "votes": 1, "replies": [ { "id": 1183853, "postDate": "2021-02-03T09:02:31.867Z", "content": "I used only the images with 1 label per image at first for training (training notebook version1).
\nThen I used the both images with 1-label and multiple-labels (training notebook version2), which improved LB:0.05.
Hi! I am new to machine learning as well as os work. I'm attempting to repurpose this and use SCnet instead of the mask rnn, when i change the config to a SCnet, i keep getting an error \"subprocess.CalledProcessError: returned non-zero exit status 1.\"
", "rawMarkdown": "Hi! I am new to machine learning as well as os work. I'm attempting to repurpose this and use SCnet instead of the mask rnn, when i change the config to a SCnet, i keep getting an error \"subprocess.CalledProcessError: returned non-zero exit status 1.\"" }, { "id": 1186522, "postDate": "2021-02-04T21:48:57.427Z", "content": "Why do you need to use instance segmentation network if CellSegmentator already does the job? I'm trying to understand the key difference between your approach and the one here. You do the instance segmentation that assigns label for each cell, and it's not multilabel, right? Can't understand why classification network on top of cropped cell patches from CellSegmentator doesn't work almost at all. It seems to me like it should be the same.
", "rawMarkdown": "Why do you need to use instance segmentation network if CellSegmentator already does the job? I'm trying to understand the key difference between your approach and the one [here](https://www.kaggle.com/c/hpa-single-cell-image-classification/discussion/216926). You do the instance segmentation that assigns label for each cell, and it's not multilabel, right? Can't understand why classification network on top of cropped cell patches from CellSegmentator doesn't work almost at all. It seems to me like it should be the same.", "replies": [ { "id": 1186821, "postDate": "2021-02-05T04:15:52.420Z", "content": "CellSegmentator can only identify which cell it is (1st cell, 2nd cell, …), but cannot get their 19-classes.
", "rawMarkdown": "CellSegmentator can only identify which cell it is (1st cell, 2nd cell, ...), but cannot get their 19-classes.", "votes": 1 }, { "id": 1187326, "postDate": "2021-02-05T11:19:51.777Z", "content": "The train.csv contrains multiple label so how do you assign label when you train a classification network on top of cropped cell patches? Or you just train single label images?
", "rawMarkdown": "The train.csv contrains multiple label so how do you assign label when you train a classification network on top of cropped cell patches? Or you just train single label images?" }, { "id": 1187363, "postDate": "2021-02-05T11:38:14.683Z", "content": "Yes, I train only single label images, assuming that all cells belong to this label (which is not correct assumption, but still should work for many cases). But it doesn't work on LB
", "rawMarkdown": "Yes, I train only single label images, assuming that all cells belong to this label (which is not correct assumption, but still should work for many cases). But it doesn't work on LB" }, { "id": 1187584, "postDate": "2021-02-05T14:50:22.740Z", "content": "The task of this competition is (weakly supervised) instance segmentation, so I thought it is a natural thing to use mmdetection mask rcnn model as a my first step.
\nI don't know what caused the big difference in score with @dschettler8845's kernel, but I think that two-stage model with CellSegmentator as a first stage will work in this competition.
Hi! Could u clarify on cell labeling please? Do u take into account only the images with 1 label per image? What about the images with multiple labels? Thanks!
", "votes": 1, "replies": [ { "id": 1183853, "author_name": "tito", "author_url": "", "post_date": "2021-02-03T09:02:31.867000", "content": "I used only the images with 1 label per image at first for training (training notebook version1).
\nThen I used the both images with 1-label and multiple-labels (training notebook version2), which improved LB:0.05.
Hi! I am new to machine learning as well as os work. I'm attempting to repurpose this and use SCnet instead of the mask rnn, when i change the config to a SCnet, i keep getting an error \"subprocess.CalledProcessError: returned non-zero exit status 1.\"
", "votes": 0, "replies": [] }, { "id": 1186522, "author_name": "Vladislav Ostankovich", "author_url": "", "post_date": "2021-02-04T21:48:57.427000", "content": "Why do you need to use instance segmentation network if CellSegmentator already does the job? I'm trying to understand the key difference between your approach and the one here. You do the instance segmentation that assigns label for each cell, and it's not multilabel, right? Can't understand why classification network on top of cropped cell patches from CellSegmentator doesn't work almost at all. It seems to me like it should be the same.
", "votes": 0, "replies": [ { "id": 1186821, "author_name": "Leon", "author_url": "", "post_date": "2021-02-05T04:15:52.420000", "content": "CellSegmentator can only identify which cell it is (1st cell, 2nd cell, …), but cannot get their 19-classes.
", "votes": 1, "replies": [] }, { "id": 1187326, "author_name": "Linye Li", "author_url": "", "post_date": "2021-02-05T11:19:51.777000", "content": "The train.csv contrains multiple label so how do you assign label when you train a classification network on top of cropped cell patches? Or you just train single label images?
", "votes": 0, "replies": [] }, { "id": 1187363, "author_name": "Vladislav Ostankovich", "author_url": "", "post_date": "2021-02-05T11:38:14.683000", "content": "Yes, I train only single label images, assuming that all cells belong to this label (which is not correct assumption, but still should work for many cases). But it doesn't work on LB
", "votes": 0, "replies": [] }, { "id": 1187584, "author_name": "tito", "author_url": "", "post_date": "2021-02-05T14:50:22.740000", "content": "The task of this competition is (weakly supervised) instance segmentation, so I thought it is a natural thing to use mmdetection mask rcnn model as a my first step.
\nI don't know what caused the big difference in score with @dschettler8845's kernel, but I think that two-stage model with CellSegmentator as a first stage will work in this competition.
\nModel\n mask rcnn\nExtract cells\n I segmented cell masks using CellSegmentator.\nCell Labeling\n labels for each image was assigned to all segmented cells in that image.\nInput colors\n only used RGB, Y is ignored here.\nBackbone\n resnest101\nImage size\n multiscale training 1024-1333\n 1280 for inference\nAugmentation & TTA\n Horizontal FLIP\nScore\n 0.288 on the current macro F1 LB, using 10K (half) training images.\n (I'll train with full data later)\n\n\nThey are both simple notebooks, so they may be useful for beginners of Instance Segmentation.\nBy changing the mmdetection configuration file, you can easily experiment with different models.\n\nHave fan!", "1183254": "Hi! Could u clarify on cell labeling please? Do u take into account only the images with 1 label per image? What about the images with multiple labels? Thanks!", "1284459": "Hi! I am new to machine learning as well as os work. I'm attempting to repurpose this and use SCnet instead of the mask rnn, when i change the config to a SCnet, i keep getting an error \"subprocess.CalledProcessError: returned non-zero exit status 1.\"", "1186522": "Why do you need to use instance segmentation network if CellSegmentator already does the job? I'm trying to understand the key difference between your approach and the one [here](https://www.kaggle.com/c/hpa-single-cell-image-classification/discussion/216926). You do the instance segmentation that assigns label for each cell, and it's not multilabel, right? Can't understand why classification network on top of cropped cell patches from CellSegmentator doesn't work almost at all. It seems to me like it should be the same." } }