![\"Example](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1225788%2Fa1fa592d0221ab8485202fb975513c2b%2Fpoor_annotation.JPG?generation=1656049893098565&alt=media\")
Like with last year's challenge, there are examples with inconsistently annotated FTU (functional tissue units).
\nIf we collect the bad annotations, I will manually create a manually corrected version and share it here with you.
\nDisclaimer: I am a physician, but no pathologist.
Like with last year's challenge, there are examples with inconsistently annotated FTU (functional tissue units).
\nIf we collect the bad annotations, I will manually create a manually corrected version and share it here with you.
\nDisclaimer: I am a physician, but no pathologist.
seems that annotation problems of last competition persists.
\n(prepare for some shakeup :( )
(i check both csv rle and json polygon below. But kagglers may want to check themselves in case i make mistakes about (1,1) pixel offset in rle)
\n![\"https://i.ibb.co/cY0Y15z/Selection-021.png\"](\"https://i.ibb.co/cY0Y15z/Selection-021.png\")
but maybe we don't have to care so much because private test (hubmap)set is totally different domain.
\nnote:
\none may want to refer to:
\nhttps://www.kaggle.com/c/hubmap-kidney-segmentation/discussion/238024
\nhttps://www.kaggle.com/code/theoviel/hubmap-final-methodology-submission/notebook
We should also compare all the csv rle annotations with json polygon annotations. Are they all identical? Or are there cases where json differ from rle?
", "rawMarkdown": "We should also compare all the csv rle annotations with json polygon annotations. Are they all identical? Or are there cases where json differ from rle?", "votes": null }, { "id": "1835329", "postDate": "06/27/2022 16:54:56", "content": "The json versions are more raw. The annotations might have issues that don't exist in the RLE copies, like overlaps, but can also allow you to distinguish between multiple adjacent FTUs which would all end up in the same mask with RLE.
", "rawMarkdown": "The json versions are more raw. The annotations might have issues that don't exist in the RLE copies, like overlaps, but can also allow you to distinguish between multiple adjacent FTUs which would all end up in the same mask with RLE.", "votes": null }, { "id": "1839006", "postDate": "07/01/2022 03:49:48", "content": "Hmm…
\nMissing annotations are kinda a problem for learning AI since he can be confused about which stuff is the real stuff…
\nI would joust split images on tiles and take tiles that have masks pixels on.
I would really like a comment from the contributors @sohier if they can comment on why this is not annotated.
", "rawMarkdown": "I would really like a comment from the contributors @sohier if they can comment on why this is not annotated.", "votes": null }, { "id": "1851810", "postDate": "07/11/2022 14:50:56", "content": "I manually checked them and they were pretty much the same.
", "rawMarkdown": "I manually checked them and they were pretty much the same.", "votes": null }, { "id": "1851812", "postDate": "07/11/2022 14:52:43", "content": "Kaggle competitions have this ambiguous rule.
\nSubmissions may not use or incorporate information from hand labeling or human prediction of the validation dataset or test data records.
I wonder if it's okay to correct those annotations manually. Even if we do it, there is no way to detect it.
", "rawMarkdown": "Kaggle competitions have this ambiguous rule.\n\n`Submissions may not use or incorporate information from hand labeling or human prediction of the validation dataset or test data records.`\n\nI wonder if it's okay to correct those annotations manually. Even if we do it, there is no way to detect it.", "votes": null }, { "id": "1852118", "postDate": "07/11/2022 19:29:06", "content": "In the last competition, it was allowed to fix the annotations, as long, as everyone got the labels.
\nI am currently not sure, if they would be of any use, and I am focussing on the training script.
\nMaybe the Lung category would benefit the most from proper annotations, but we don't know how the validation/test data is annotated.
Yeah, but most of the time everyone got the labels just before the data disclose deadline. Models are good at estimating noise free distribution from this amount of noisy labels but it would definitely be useful for validation purposes.
", "rawMarkdown": "Yeah, but most of the time everyone got the labels just before the data disclose deadline. Models are good at estimating noise free distribution from this amount of noisy labels but it would definitely be useful for validation purposes.", "votes": null }, { "id": "1860815", "postDate": "07/18/2022 15:29:41", "content": "![\"IMG](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1225788%2F178b6496a1eb84b001a8545d54ee2aa4%2Fpred%204944.png?generation=1658158022517544&alt=media\")
Image taken from the validation split.
\nRed is false positive, green true positive, blue false negative.
\nYou can see, that only a few of the glands are annotated. These \"annotation errors\" makes cv very hard, since a good model will have a poor score in this images and other poorly annotated ones.
from my understanding looking at the images shared, and the orientation issue with annotation co-ordinates.
\nIt simply calls for a preprocessing step - Align, align the tissue samples with the json or rle coords provided.
\nAlso since two separate data for annotation coordinates are given i think, just an intuition honestly, that difference between the the two annotation might give clue to rotation need to align tissue sample with annotation.
seems that annotation problems of last competition persists.
\n(prepare for some shakeup :( )
(i check both csv rle and json polygon below. But kagglers may want to check themselves in case i make mistakes about (1,1) pixel offset in rle)
\n
but maybe we don't have to care so much because private test (hubmap)set is totally different domain.
\nnote:
\none may want to refer to:
\nhttps://www.kaggle.com/c/hubmap-kidney-segmentation/discussion/238024
\nhttps://www.kaggle.com/code/theoviel/hubmap-final-methodology-submission/notebook
We should also compare all the csv rle annotations with json polygon annotations. Are they all identical? Or are there cases where json differ from rle?
", "votes": null, "replies": [ { "id": 1835329, "author_name": "sohier", "author_url": "", "post_date": "06/27/2022 16:54:56", "content": "The json versions are more raw. The annotations might have issues that don't exist in the RLE copies, like overlaps, but can also allow you to distinguish between multiple adjacent FTUs which would all end up in the same mask with RLE.
", "votes": null, "replies": [] }, { "id": 1851810, "author_name": "gunesevitan", "author_url": "", "post_date": "07/11/2022 14:50:56", "content": "I manually checked them and they were pretty much the same.
", "votes": null, "replies": [] } ] }, { "id": 1839006, "author_name": "urosjarc", "author_url": "", "post_date": "07/01/2022 03:49:48", "content": "Hmm…
\nMissing annotations are kinda a problem for learning AI since he can be confused about which stuff is the real stuff…
\nI would joust split images on tiles and take tiles that have masks pixels on.
I would really like a comment from the contributors @sohier if they can comment on why this is not annotated.
", "votes": null, "replies": [] }, { "id": 1851812, "author_name": "gunesevitan", "author_url": "", "post_date": "07/11/2022 14:52:43", "content": "Kaggle competitions have this ambiguous rule.
\nSubmissions may not use or incorporate information from hand labeling or human prediction of the validation dataset or test data records.
I wonder if it's okay to correct those annotations manually. Even if we do it, there is no way to detect it.
", "votes": null, "replies": [ { "id": 1852118, "author_name": "theudas", "author_url": "", "post_date": "07/11/2022 19:29:06", "content": "In the last competition, it was allowed to fix the annotations, as long, as everyone got the labels.
\nI am currently not sure, if they would be of any use, and I am focussing on the training script.
\nMaybe the Lung category would benefit the most from proper annotations, but we don't know how the validation/test data is annotated.
Yeah, but most of the time everyone got the labels just before the data disclose deadline. Models are good at estimating noise free distribution from this amount of noisy labels but it would definitely be useful for validation purposes.
", "votes": null, "replies": [] } ] }, { "id": 1860815, "author_name": "theudas", "author_url": "", "post_date": "07/18/2022 15:29:41", "content": "
Image taken from the validation split.
\nRed is false positive, green true positive, blue false negative.
\nYou can see, that only a few of the glands are annotated. These \"annotation errors\" makes cv very hard, since a good model will have a poor score in this images and other poorly annotated ones.
from my understanding looking at the images shared, and the orientation issue with annotation co-ordinates.
\nIt simply calls for a preprocessing step - Align, align the tissue samples with the json or rle coords provided.
\nAlso since two separate data for annotation coordinates are given i think, just an intuition honestly, that difference between the the two annotation might give clue to rotation need to align tissue sample with annotation.