If a picture with dark areas like afa5e8098.tiff is in private data, can the model predict well?
\n--> Of course, using augmentation can solve this problem.
But what if annotation is ambiguous?
\nLet's check the image below.(afa5e8098.tiff)
\nIn the image below, green is the prediction mask and red is the ground truth.
https://www.kaggle.com/gwanghan/discussion-images?select=predict.PNG
\nhttps://www.kaggle.com/gwanghan/discussion-images?select=raw.PNG
Some of the glomeruli in the dark are not annotated as expected (at least to my eyes they appear to be ordinary glomeruli).
\nAlso, dark areas of the image other than the link image appear glomeruli, but are not annotated.
I know that there are noise and bayes errors in the dataset in any competition. However, there are a lot of issues with annotation, especially in this competition.
\nSince private data can also have labeling issues like this, it is too difficult for me how to make the model robust.
", "messages": [ { "id": 1274617, "postDate": "2021-04-15T13:25:01.033Z", "content": "Roll the dice 🎲
", "rawMarkdown": "Roll the dice 🎲", "votes": 3 }, { "id": 1274271, "postDate": "2021-04-15T06:45:41.430Z", "content": "If a picture with dark areas like afa5e8098.tiff is in private data, can the model predict well?
\n--> Of course, using augmentation can solve this problem.
But what if annotation is ambiguous?
\nLet's check the image below.(afa5e8098.tiff)
\nIn the image below, green is the prediction mask and red is the ground truth.
https://www.kaggle.com/gwanghan/discussion-images?select=predict.PNG
\nhttps://www.kaggle.com/gwanghan/discussion-images?select=raw.PNG
Some of the glomeruli in the dark are not annotated as expected (at least to my eyes they appear to be ordinary glomeruli).
\nAlso, dark areas of the image other than the link image appear glomeruli, but are not annotated.
I know that there are noise and bayes errors in the dataset in any competition. However, there are a lot of issues with annotation, especially in this competition.
\nSince private data can also have labeling issues like this, it is too difficult for me how to make the model robust.
", "rawMarkdown": "If a picture with dark areas like **afa5e8098.tiff** is in private data, can the model predict well?\n--> Of course, using **augmentation** can solve this problem.\n\nBut what if **annotation** is **ambiguous**?\n\nLet's check the image below.(**afa5e8098.tiff**)\nIn the image below, **green** is the **prediction mask** and **red** is the **ground truth**.\n\nhttps://www.kaggle.com/gwanghan/discussion-images?select=predict.PNG\nhttps://www.kaggle.com/gwanghan/discussion-images?select=raw.PNG\n\nSome of the glomeruli in the dark are not annotated as expected (at least to my eyes they appear to be ordinary glomeruli).\nAlso, dark areas of the image other than the link image appear glomeruli, but are not annotated.\n\nI know that there are **noise** and **bayes errors** in the dataset in **any competition**. However, there are a lot of issues with annotation, especially in this competition.\n\n*****************\n\nSince private data can also have labeling issues like this, it is too difficult for me how to make the model robust.\n", "votes": 4 }, { "id": 1274713, "postDate": "2021-04-15T14:43:36.293Z", "content": "What kind of augmentation do you use to improve dark image performance if you don't mind me asking?
", "rawMarkdown": "What kind of augmentation do you use to improve dark image performance if you don't mind me asking?", "votes": 1, "replies": [ { "id": 1275160, "postDate": "2021-04-16T02:39:24.617Z", "content": "It was helpful when I used HueSaturationValue augmentation :)
", "rawMarkdown": "It was helpful when I used HueSaturationValue augmentation :)" }, { "id": 1277773, "postDate": "2021-04-19T07:40:26.633Z", "content": "I want to ask where you use this augmentation? Training or predicting? Just because I have tried it in training but get lower LB :(
\nOf course, if you don't want to reveal too many details before the end of the game, I can understand.
You should use this in training and also LB means nothing in this comp
", "rawMarkdown": "You should use this in training and also LB means nothing in this comp" }, { "id": 1278244, "postDate": "2021-04-19T17:23:44.647Z", "content": "@chinesewuji I used it for training
", "rawMarkdown": "@chinesewuji I used it for training", "votes": 1 }, { "id": 1278525, "postDate": "2021-04-20T02:21:42.030Z", "content": "Thanks for your sharing. I will try more to imporve the robustness of my mdoel.
", "rawMarkdown": "Thanks for your sharing. I will try more to imporve the robustness of my mdoel." } ] }, { "id": 1274566, "postDate": "2021-04-15T12:16:26.460Z", "content": "Hi @gwanghan The links to your images don't work for me? User error? or are they not included properly?
\nI'am curious to see your results.
Sorry, I changed it to public :)
", "rawMarkdown": "Sorry, I changed it to public :)" } ] }, { "id": 1274320, "postDate": "2021-04-15T07:27:45.693Z", "content": "I agreed. As my last Disscussion has said. I found there are lots of noises in afa5e8098 by classifier. https://www.kaggle.com/c/hubmap-kidney-segmentation/discussion/232097 @katyborner
", "rawMarkdown": "I agreed. As my last Disscussion has said. I found there are lots of noises in afa5e8098 by classifier. https://www.kaggle.com/c/hubmap-kidney-segmentation/discussion/232097 @katyborner ", "votes": 1, "replies": [ { "id": 1277025, "postDate": "2021-04-18T09:49:31.843Z", "rawMarkdown": "", "votes": 2, "isDeleted": true }, { "id": 1277859, "postDate": "2021-04-19T09:56:18.447Z", "content": "Hi @narainp can you share your thoughts why dice varies so much among sildes? maybe noise annotation?
", "rawMarkdown": "Hi @narainp can you share your thoughts why dice varies so much among sildes? maybe noise annotation?" }, { "id": 1279308, "postDate": "2021-04-20T19:17:07.820Z", "content": "Very good!
", "rawMarkdown": "Very good!" }, { "id": 1280805, "postDate": "2021-04-22T11:24:11.110Z", "rawMarkdown": "", "isDeleted": true } ] } ], "comments": [ { "id": 1274617, "author_name": "cool_rabbit", "author_url": "", "post_date": "2021-04-15T13:25:01.033000", "content": "Roll the dice 🎲
", "votes": 3, "replies": [] }, { "id": 1274713, "author_name": "Shujun", "author_url": "", "post_date": "2021-04-15T14:43:36.293000", "content": "What kind of augmentation do you use to improve dark image performance if you don't mind me asking?
", "votes": 1, "replies": [ { "id": 1275160, "author_name": "starsnew", "author_url": "", "post_date": "2021-04-16T02:39:24.617000", "content": "It was helpful when I used HueSaturationValue augmentation :)
", "votes": 0, "replies": [] }, { "id": 1277773, "author_name": "Matthew Wu", "author_url": "", "post_date": "2021-04-19T07:40:26.633000", "content": "I want to ask where you use this augmentation? Training or predicting? Just because I have tried it in training but get lower LB :(
\nOf course, if you don't want to reveal too many details before the end of the game, I can understand.
You should use this in training and also LB means nothing in this comp
", "votes": 0, "replies": [] }, { "id": 1278244, "author_name": "starsnew", "author_url": "", "post_date": "2021-04-19T17:23:44.647000", "content": "@chinesewuji I used it for training
", "votes": 1, "replies": [] }, { "id": 1278525, "author_name": "Matthew Wu", "author_url": "", "post_date": "2021-04-20T02:21:42.030000", "content": "Thanks for your sharing. I will try more to imporve the robustness of my mdoel.
", "votes": 0, "replies": [] } ] }, { "id": 1274566, "author_name": "Robin Smits", "author_url": "", "post_date": "2021-04-15T12:16:26.460000", "content": "Hi @gwanghan The links to your images don't work for me? User error? or are they not included properly?
\nI'am curious to see your results.
Sorry, I changed it to public :)
", "votes": 0, "replies": [] } ] }, { "id": 1274320, "author_name": "朴大福", "author_url": "", "post_date": "2021-04-15T07:27:45.693000", "content": "I agreed. As my last Disscussion has said. I found there are lots of noises in afa5e8098 by classifier. https://www.kaggle.com/c/hubmap-kidney-segmentation/discussion/232097 @katyborner
", "votes": 1, "replies": [ { "id": 1277025, "author_name": "", "author_url": "", "post_date": "2021-04-18T09:49:31.843000", "content": "", "votes": 2, "replies": [] }, { "id": 1277859, "author_name": "Yi Wu", "author_url": "", "post_date": "2021-04-19T09:56:18.447000", "content": "Hi @narainp can you share your thoughts why dice varies so much among sildes? maybe noise annotation?
", "votes": 0, "replies": [] }, { "id": 1279308, "author_name": "victorasso", "author_url": "", "post_date": "2021-04-20T19:17:07.820000", "content": "Very good!
", "votes": 0, "replies": [] }, { "id": 1280805, "author_name": "", "author_url": "", "post_date": "2021-04-22T11:24:11.110000", "content": "", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "1274617": "Roll the dice 🎲", "1274271": "If a picture with dark areas like **afa5e8098.tiff** is in private data, can the model predict well?\n--> Of course, using **augmentation** can solve this problem.\n\nBut what if **annotation** is **ambiguous**?\n\nLet's check the image below.(**afa5e8098.tiff**)\nIn the image below, **green** is the **prediction mask** and **red** is the **ground truth**.\n\nhttps://www.kaggle.com/gwanghan/discussion-images?select=predict.PNG\nhttps://www.kaggle.com/gwanghan/discussion-images?select=raw.PNG\n\nSome of the glomeruli in the dark are not annotated as expected (at least to my eyes they appear to be ordinary glomeruli).\nAlso, dark areas of the image other than the link image appear glomeruli, but are not annotated.\n\nI know that there are **noise** and **bayes errors** in the dataset in **any competition**. However, there are a lot of issues with annotation, especially in this competition.\n\n*****************\n\nSince private data can also have labeling issues like this, it is too difficult for me how to make the model robust.\n", "1274713": "What kind of augmentation do you use to improve dark image performance if you don't mind me asking?", "1274566": "Hi @gwanghan The links to your images don't work for me? User error? or are they not included properly?\nI'am curious to see your results.", "1274320": "I agreed. As my last Disscussion has said. I found there are lots of noises in afa5e8098 by classifier. https://www.kaggle.com/c/hubmap-kidney-segmentation/discussion/232097 @katyborner " } }