![\"aa05346ff-10752-11520-12288-13056\"](\"https://i.postimg.cc/QBvnyxwR/aa05346ff-10752-11520-12288-13056.png\")
![\"aa05346ff-10752-11520-16128-16896\"](\"https://i.postimg.cc/RWV83Ywb/aa05346ff-10752-11520-16128-16896.png\")
![\"aa05346ff-10752-11520-3072-3840\"](\"https://i.postimg.cc/sGY8Sdnb/aa05346ff-10752-11520-3072-3840.png\")
\n
![\"aa05346ff-11520-12288-16128-16896\"](\"https://i.postimg.cc/yDQG0TP6/aa05346ff-11520-12288-16128-16896.png\")
![\"aa05346ff-11520-12288-19200-19968\"](\"https://i.postimg.cc/KR5w8b3M/aa05346ff-11520-12288-19200-19968.png\")
![\"aa05346ff-11520-12288-9216-9984\"](\"https://i.postimg.cc/jWDFcJgH/aa05346ff-11520-12288-9216-9984.png\")
\n
![\"aa05346ff-12288-13056-2304-3072\"](\"https://i.postimg.cc/JG8S8008/aa05346ff-12288-13056-2304-3072.png\")
Are the competition sponsors aware that there are significant blurriness and exposure issues with image aa05346ff.tiff from the test set? It's not noticeable from the thumbnails and resized images. But in its true size, large portions of the image are very blurry and lack detail.
This doesn't impact the training part of the competition. But it might make the competition seem less successful than it actually is in terms of producing a viable model. Especially if every model's final accuracy on the test dataset is 5-10% lower because of one test image.
\nThe following are full resolution patches with no modifications on my part.
\n
\n
\n
Are the competition sponsors aware that there are significant blurriness and exposure issues with image aa05346ff.tiff from the test set? It's not noticeable from the thumbnails and resized images. But in its true size, large portions of the image are very blurry and lack detail.
This doesn't impact the training part of the competition. But it might make the competition seem less successful than it actually is in terms of producing a viable model. Especially if every model's final accuracy on the test dataset is 5-10% lower because of one test image.
\nThe following are full resolution patches with no modifications on my part.
\n
\n
\n
",
"votes": null
},
{
"id": "1251044",
"postDate": "03/24/2021 13:08:54",
"content": "I think this is part of the task, to produce good results even when artifacts are present.
\nI am not sure whether these artifact are due to errors in sample processing or dues to post processing by the organizers. The test images from the original public test set that had artifacts (c68f and afa5) where released on hubmap with the same artifacts they had in the competition tiffs. So my guess is that these are due to errors in processing, handling, scanning, etc of the samples
\nIMO, the ultimate challenge of this competition is not only to get a good performance generalized model but that it also needs to perform well on images with artifacts it possibly hasn`t seen before. Who knows what waits for un in the private testset
", "rawMarkdown": "I think this is part of the task, to produce good results even when artifacts are present.\n\nI am not sure whether these artifact are due to errors in sample processing or dues to post processing by the organizers. The test images from the original public test set that had artifacts (c68f and afa5) where released on hubmap with the same artifacts they had in the competition tiffs. So my guess is that these are due to errors in processing, handling, scanning, etc of the samples\n\nIMO, the ultimate challenge of this competition is not only to get a good performance generalized model but that it also needs to perform well on images with artifacts it possibly hasn`t seen before. Who knows what waits for un in the private testset", "votes": null }, { "id": "1251674", "postDate": "03/25/2021 02:36:48", "content": "I worked for a hospital in the pathology department and creating histology images can be hard. Most of the scans we did had some blurr on the edges and some kind of rectangular artifacts like you show here so i think this is normal and you can't really get better quality.
", "rawMarkdown": "I worked for a hospital in the pathology department and creating histology images can be hard. Most of the scans we did had some blurr on the edges and some kind of rectangular artifacts like you show here so i think this is normal and you can't really get better quality.", "votes": null } ], "comments": [ { "id": 1251044, "author_name": "rosuluc", "author_url": "", "post_date": "03/24/2021 13:08:54", "content": "I think this is part of the task, to produce good results even when artifacts are present.
\nI am not sure whether these artifact are due to errors in sample processing or dues to post processing by the organizers. The test images from the original public test set that had artifacts (c68f and afa5) where released on hubmap with the same artifacts they had in the competition tiffs. So my guess is that these are due to errors in processing, handling, scanning, etc of the samples
\nIMO, the ultimate challenge of this competition is not only to get a good performance generalized model but that it also needs to perform well on images with artifacts it possibly hasn`t seen before. Who knows what waits for un in the private testset
", "votes": null, "replies": [] }, { "id": 1251674, "author_name": "yannmajewski", "author_url": "", "post_date": "03/25/2021 02:36:48", "content": "I worked for a hospital in the pathology department and creating histology images can be hard. Most of the scans we did had some blurr on the edges and some kind of rectangular artifacts like you show here so i think this is normal and you can't really get better quality.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1248960": "Are the competition sponsors aware that there are significant blurriness and exposure issues with image `aa05346ff.tiff` from the test set? It's not noticeable from the thumbnails and resized images. But in its true size, large portions of the image are very blurry and lack detail.\n\nThis doesn't impact the training part of the competition. But it might make the competition seem less successful than it actually is in terms of producing a viable model. Especially if every model's final accuracy on the test dataset is 5-10% lower because of one test image.\n\nThe following are full resolution patches with no modifications on my part.\n\n ",
"1251044": "I think this is part of the task, to produce good results even when artifacts are present.\n\nI am not sure whether these artifact are due to errors in sample processing or dues to post processing by the organizers. The test images from the original public test set that had artifacts (c68f and afa5) where released on hubmap with the same artifacts they had in the competition tiffs. So my guess is that these are due to errors in processing, handling, scanning, etc of the samples\n\nIMO, the ultimate challenge of this competition is not only to get a good performance generalized model but that it also needs to perform well on images with artifacts it possibly hasn`t seen before. Who knows what waits for un in the private testset",
"1251674": "I worked for a hospital in the pathology department and creating histology images can be hard. Most of the scans we did had some blurr on the edges and some kind of rectangular artifacts like you show here so i think this is normal and you can't really get better quality."
},
"source": "meta"
}