\n
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F8251891%2Fc7e4ca43c5f5418583983bbb1ec0a08c%2F(1).png?generation=1703326345488635&alt=media\")
It seems there might be an issue when applying size-scale transformations (such as resize, shiftscalerotate, etc.) using the albumentations library(perhaps due to a problem in quantization process)
\nIn the attached images, the left line shows a mask resized through the \"albumentations.Resize\", while the right line exhibits a mask resized using \"cv2.resize\" without quantization. (Those pictures are the result of reducing the original size by a factor of x4.)
\n
The left mask was obtained using the code:
\nA.Resize(h, w)(=img, =mask)\nWhile the right mask was obtained using the code:
\ncv2(, (w, h))\ncv2(, (w, h))\nCould it be that I've used the code incorrectly?
", "messages": [ { "id": 2571547, "postDate": "2023-12-23T10:16:29.377Z", "content": "It seems there might be an issue when applying size-scale transformations (such as resize, shiftscalerotate, etc.) using the albumentations library(perhaps due to a problem in quantization process)
\nIn the attached images, the left line shows a mask resized through the \"albumentations.Resize\", while the right line exhibits a mask resized using \"cv2.resize\" without quantization. (Those pictures are the result of reducing the original size by a factor of x4.)
\n
The left mask was obtained using the code:
\nA.Resize(h, w)(=img, =mask)\nWhile the right mask was obtained using the code:
\ncv2(, (w, h))\ncv2(, (w, h))\nCould it be that I've used the code incorrectly?
", "rawMarkdown": "It seems there might be an issue when applying size-scale transformations (such as resize, shiftscalerotate, etc.) using the albumentations library(perhaps due to a problem in quantization process)\n\nIn the attached images, the left line shows a mask resized through the \"albumentations.Resize\", while the right line exhibits a mask resized using \"cv2.resize\" without quantization. (Those pictures are the result of reducing the original size by a factor of x4.) \n.png?generation=1703326345488635&alt=media)\n\nThe left mask was obtained using the code:\n\n A.Resize(h, w)(image=img, mask=mask)\n\nWhile the right mask was obtained using the code: \n \n cv2.resize(img, (w, h))\n cv2.resize(mask, (w, h))\n\nCould it be that I've used the code incorrectly?", "votes": 9 }, { "id": 2571999, "postDate": "2023-12-23T19:22:01.220Z", "content": "Resizing is the key in winning the private dataset.
", "rawMarkdown": "Resizing is the key in winning the private dataset.", "votes": 3, "replies": [ { "id": 2572202, "postDate": "2023-12-24T02:10:11.510Z", "content": "I really think so too!
\nIn this competition, precisely segmenting small details is crucial.
No. Private test data is of different resolution. How to use resize correctly to ensure your model is robust against resolution is the key
", "rawMarkdown": "No. Private test data is of different resolution. How to use resize correctly to ensure your model is robust against resolution is the key", "votes": 5 }, { "id": 2572697, "postDate": "2023-12-24T13:18:32.680Z", "content": "Ah, I see. So you meant “resize” in that sense. Indeed, creating a model that's robust to resolution could significantly impact achieving good results in private. Thank you for the valuable insight.
", "rawMarkdown": "Ah, I see. So you meant “resize” in that sense. Indeed, creating a model that's robust to resolution could significantly impact achieving good results in private. Thank you for the valuable insight." } ] } ] }, { "id": 2571681, "postDate": "2023-12-23T12:58:16.137Z", "content": "I think the interpolation methods are different. For Albumentations' resize applied to masks, cv2.INTER_NEAREST is used.
", "rawMarkdown": "I think the interpolation methods are different. For Albumentations' resize applied to masks, cv2.INTER_NEAREST is used.", "votes": 3, "replies": [ { "id": 2571699, "postDate": "2023-12-23T13:24:12.807Z", "content": "Thank you!
\nSetting it as cv2.resize(mask, (w, h), interpolation=cv2.INTER_NEAREST) produced the same result!
Did you correctly set the arguments?
\nIs it possible that you forgot \"interpolation=\"?
\ncv2.resize(mask, (w, h), interpolation = cv2.INTER_NEAREST) is correct.
I meant to say that your statement was correct - A.Resize and cv2.resize produce the same result.
\nThanks!😃
I misunderstood. Glad to hear it's resolved!
", "rawMarkdown": "I misunderstood. Glad to hear it's resolved!" } ] } ] }, { "id": 2571585, "postDate": "2023-12-23T10:56:19.790Z", "content": "cv2.resize expects (w, h) :
\n cv2.resize(img, (w, h) )
Ah, right! Thank you. 😀
\nHowever, in this case where h=w=16, it still seems there is an issue.
Hmm, then that's weird because Albumentations uses cv2 under the hood, and both employ bilinear interpolation by default.
", "rawMarkdown": "Hmm, then that's weird because Albumentations uses cv2 under the hood, and both employ bilinear interpolation by default." } ] } ] }, { "id": 2573355, "postDate": "2023-12-24T23:04:01.007Z", "rawMarkdown": "", "isDeleted": true, "replies": [ { "id": 2573421, "postDate": "2023-12-25T03:27:35.810Z", "content": "When I tried augmenting with the mask resolution(Resize, ShiftScaleRotate, …), I found through several experiments that it actually worsened the performance.
\nI think it’s because of using cv2.INTER_NEAREST for mask interpolation in the Albumentations library.
\nIndeed, using interpolation=cv2.INTER_LINEAR for mask yielded better results.
", "rawMarkdown": "When I tried augmenting with the mask resolution(Resize, ShiftScaleRotate, …), I found through several experiments that it actually worsened the performance.\n\nI think it’s because of using cv2.INTER_NEAREST for mask interpolation in the Albumentations library.\n\nIndeed, using interpolation=cv2.INTER_LINEAR for mask yielded better results.", "votes": 3, "replies": [ { "id": 2573424, "postDate": "2023-12-25T03:36:00.023Z", "content": "Noted! I will share if I find anything helpful in further research
", "rawMarkdown": "Noted! I will share if I find anything helpful in further research ", "votes": 2 } ] }, { "id": 2573649, "postDate": "2023-12-25T08:49:03.420Z", "content": "What resolution do you use to train the model?
", "rawMarkdown": "What resolution do you use to train the model?" } ] } ], "comments": [ { "id": 2571999, "author_name": "hengck23", "author_url": "", "post_date": "2023-12-23T19:22:01.220000", "content": "Resizing is the key in winning the private dataset.
", "votes": 3, "replies": [ { "id": 2572202, "author_name": "siwooyong", "author_url": "", "post_date": "2023-12-24T02:10:11.510000", "content": "I really think so too!
\nIn this competition, precisely segmenting small details is crucial.
No. Private test data is of different resolution. How to use resize correctly to ensure your model is robust against resolution is the key
", "votes": 5, "replies": [] }, { "id": 2572697, "author_name": "siwooyong", "author_url": "", "post_date": "2023-12-24T13:18:32.680000", "content": "Ah, I see. So you meant “resize” in that sense. Indeed, creating a model that's robust to resolution could significantly impact achieving good results in private. Thank you for the valuable insight.
", "votes": 0, "replies": [] } ] } ] }, { "id": 2571681, "author_name": "Takayuki Higuchi", "author_url": "", "post_date": "2023-12-23T12:58:16.137000", "content": "I think the interpolation methods are different. For Albumentations' resize applied to masks, cv2.INTER_NEAREST is used.
", "votes": 3, "replies": [ { "id": 2571699, "author_name": "siwooyong", "author_url": "", "post_date": "2023-12-23T13:24:12.807000", "content": "Thank you!
\nSetting it as cv2.resize(mask, (w, h), interpolation=cv2.INTER_NEAREST) produced the same result!
Did you correctly set the arguments?
\nIs it possible that you forgot \"interpolation=\"?
\ncv2.resize(mask, (w, h), interpolation = cv2.INTER_NEAREST) is correct.
I meant to say that your statement was correct - A.Resize and cv2.resize produce the same result.
\nThanks!😃
I misunderstood. Glad to hear it's resolved!
", "votes": 0, "replies": [] } ] } ] }, { "id": 2571585, "author_name": "Reacher", "author_url": "", "post_date": "2023-12-23T10:56:19.790000", "content": "cv2.resize expects (w, h) :
\n cv2.resize(img, (w, h) )
Ah, right! Thank you. 😀
\nHowever, in this case where h=w=16, it still seems there is an issue.
Hmm, then that's weird because Albumentations uses cv2 under the hood, and both employ bilinear interpolation by default.
", "votes": 0, "replies": [] } ] } ] }, { "id": 2573355, "author_name": "", "author_url": "", "post_date": "2023-12-24T23:04:01.007000", "content": "", "votes": 0, "replies": [ { "id": 2573421, "author_name": "siwooyong", "author_url": "", "post_date": "2023-12-25T03:27:35.810000", "content": "When I tried augmenting with the mask resolution(Resize, ShiftScaleRotate, …), I found through several experiments that it actually worsened the performance.
\nI think it’s because of using cv2.INTER_NEAREST for mask interpolation in the Albumentations library.
\nIndeed, using interpolation=cv2.INTER_LINEAR for mask yielded better results.
", "votes": 3, "replies": [ { "id": 2573424, "author_name": "Cody_Null", "author_url": "", "post_date": "2023-12-25T03:36:00.023000", "content": "Noted! I will share if I find anything helpful in further research
", "votes": 2, "replies": [] } ] }, { "id": 2573649, "author_name": "Arunodhayan", "author_url": "", "post_date": "2023-12-25T08:49:03.420000", "content": "What resolution do you use to train the model?
", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "2571547": "It seems there might be an issue when applying size-scale transformations (such as resize, shiftscalerotate, etc.) using the albumentations library(perhaps due to a problem in quantization process)\n\nIn the attached images, the left line shows a mask resized through the \"albumentations.Resize\", while the right line exhibits a mask resized using \"cv2.resize\" without quantization. (Those pictures are the result of reducing the original size by a factor of x4.) \n.png?generation=1703326345488635&alt=media)\n\nThe left mask was obtained using the code:\n\n A.Resize(h, w)(image=img, mask=mask)\n\nWhile the right mask was obtained using the code: \n \n cv2.resize(img, (w, h))\n cv2.resize(mask, (w, h))\n\nCould it be that I've used the code incorrectly?", "2571999": "Resizing is the key in winning the private dataset.", "2571681": "I think the interpolation methods are different. For Albumentations' resize applied to masks, cv2.INTER_NEAREST is used.", "2571585": "cv2.resize expects (w, h) :\n `cv2.resize(img, (w, h) )`", "2573355": "" } }