{ "id": 295603, "title": "RLE to mask bug", "url": "/competitions/sartorius-cell-instance-segmentation/discussion/295603", "author_name": "", "post_date": "2021-12-16T20:58:47.735713800Z", "votes": 16, "comment_count": 4, "views": 0, "content": "

When I look closely at the numpy mask converted from RLE, I noticed some of the cells are missed.

\n

Take '17d738f88487' for example,
\n\"\"
\nThe left is the correct one, and the right is the output of many high vote notebooks.

\n

To correct the bug, you can enumerate from number>0, like such as 10 or sth

\n
 for i, annot in enumerate(annotations,10):\n         img_mask = np.where(rle2mask(annot, shape)!=0, i, img_mask)\n
", "messages": [ { "id": "1620506", "postDate": "12/16/2021 20:58:47", "content": "

When I look closely at the numpy mask converted from RLE, I noticed some of the cells are missed.

\n

Take '17d738f88487' for example,
\n\"\"
\nThe left is the correct one, and the right is the output of many high vote notebooks.

\n

To correct the bug, you can enumerate from number>0, like such as 10 or sth

\n
 for i, annot in enumerate(annotations,10):\n         img_mask = np.where(rle2mask(annot, shape)!=0, i, img_mask)\n
", "rawMarkdown": "When I look closely at the numpy mask converted from RLE, I noticed some of the cells are missed. \n\nTake '17d738f88487' for example, \n![](https://i.ibb.co/ZM8nnLb/Screenshot-2021-12-16-145731.png)\nThe left is the correct one, and the right is the output of many high vote notebooks.\n\nTo correct the bug, you can enumerate from number>0, like such as 10 or sth\n\n```\n for i, annot in enumerate(annotations,10):\n img_mask = np.where(rle2mask(annot, shape)!=0, i, img_mask)\n```", "votes": null }, { "id": "1620663", "postDate": "12/17/2021 02:54:18", "content": "

I just looked at my annotation file for 17d738f88487 and it is not missing the above cell. Which notebook did you use? This is what I used: https://www.kaggle.com/ammarnassanalhajali/k-fold-crossvalidation-coco-dataset-generator?scriptVersionId=79370156

", "rawMarkdown": "I just looked at my annotation file for 17d738f88487 and it is not missing the above cell. Which notebook did you use? This is what I used: https://www.kaggle.com/ammarnassanalhajali/k-fold-crossvalidation-coco-dataset-generator?scriptVersionId=79370156", "votes": null }, { "id": "1620698", "postDate": "12/17/2021 04:30:19", "content": "

yeah i think this is the best one since it uses uncompressed RLE. A lot slower than the other generators but in my experiments it might be a reason for better performance. Haven't tried the one with clean astro annotations since I couldn't incorporate polygon masks loss calculation into my pipeline

", "rawMarkdown": "yeah i think this is the best one since it uses uncompressed RLE. A lot slower than the other generators but in my experiments it might be a reason for better performance. Haven't tried the one with clean astro annotations since I couldn't incorporate polygon masks loss calculation into my pipeline", "votes": null }, { "id": "1622277", "postDate": "12/18/2021 14:37:03", "content": "

I found this bug, too. it lost one example. Thank you for sharing.

", "rawMarkdown": "I found this bug, too. it lost one example. Thank you for sharing.", "votes": null }, { "id": "1624024", "postDate": "12/20/2021 13:31:22", "content": "

Could anyone figure out which notebooks are wrong?

", "rawMarkdown": "Could anyone figure out which notebooks are wrong?", "votes": null } ], "comments": [ { "id": 1620663, "author_name": "shujun717", "author_url": "", "post_date": "12/17/2021 02:54:18", "content": "

I just looked at my annotation file for 17d738f88487 and it is not missing the above cell. Which notebook did you use? This is what I used: https://www.kaggle.com/ammarnassanalhajali/k-fold-crossvalidation-coco-dataset-generator?scriptVersionId=79370156

", "votes": null, "replies": [ { "id": 1620698, "author_name": "ferlockx", "author_url": "", "post_date": "12/17/2021 04:30:19", "content": "

yeah i think this is the best one since it uses uncompressed RLE. A lot slower than the other generators but in my experiments it might be a reason for better performance. Haven't tried the one with clean astro annotations since I couldn't incorporate polygon masks loss calculation into my pipeline

", "votes": null, "replies": [] } ] }, { "id": 1622277, "author_name": "shigengtian", "author_url": "", "post_date": "12/18/2021 14:37:03", "content": "

I found this bug, too. it lost one example. Thank you for sharing.

", "votes": null, "replies": [] }, { "id": 1624024, "author_name": "chihantsai", "author_url": "", "post_date": "12/20/2021 13:31:22", "content": "

Could anyone figure out which notebooks are wrong?

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1620506": "When I look closely at the numpy mask converted from RLE, I noticed some of the cells are missed. \n\nTake '17d738f88487' for example, \n![](https://i.ibb.co/ZM8nnLb/Screenshot-2021-12-16-145731.png)\nThe left is the correct one, and the right is the output of many high vote notebooks.\n\nTo correct the bug, you can enumerate from number>0, like such as 10 or sth\n\n```\n for i, annot in enumerate(annotations,10):\n img_mask = np.where(rle2mask(annot, shape)!=0, i, img_mask)\n```", "1620663": "I just looked at my annotation file for 17d738f88487 and it is not missing the above cell. Which notebook did you use? This is what I used: https://www.kaggle.com/ammarnassanalhajali/k-fold-crossvalidation-coco-dataset-generator?scriptVersionId=79370156", "1620698": "yeah i think this is the best one since it uses uncompressed RLE. A lot slower than the other generators but in my experiments it might be a reason for better performance. Haven't tried the one with clean astro annotations since I couldn't incorporate polygon masks loss calculation into my pipeline", "1622277": "I found this bug, too. it lost one example. Thank you for sharing.", "1624024": "Could anyone figure out which notebooks are wrong?" }, "source": "meta" }