![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F18286739%2F2a2d8bd548bada27dd5cd8d84700ac73%2F2024-01-04%20%2014.13.13.png?generation=1704366838653455&alt=media\")
What does 'Sparsely segmented (about 65%)' mean?
\n\n\nkidney_2 - The whole of a kidney from another donor, at 50um resolution. Sparsely segmented (about 65%).
\n
What does 'Sparsely segmented (about 65%)' mean?
\n\n\nkidney_2 - The whole of a kidney from another donor, at 50um resolution. Sparsely segmented (about 65%).
\n
This means that just 65% of labels are present in the ground truth images.
", "rawMarkdown": "This means that just 65% of labels are present in the ground truth images.", "votes": null }, { "id": "2587357", "postDate": "01/04/2024 17:56:21", "content": "i wonder how did the annotator knows 65% are present?
\nhe must know all 100 % first to compute the percentage of label present.
\nin that case, why don't he just let us have the 100%?
", "rawMarkdown": "i wonder how did the annotator knows 65% are present?\n\nhe must know all 100 % first to compute the percentage of label present.\n\nin that case, why don't he just let us have the 100%?", "votes": null }, { "id": "2587380", "postDate": "01/04/2024 18:05:59", "content": "Please see: https://www.kaggle.com/competitions/blood-vessel-segmentation/discussion/454353
\nHope this helps!
@yashvrdnjain Thanks! I missed that early post.
\n\n\" five randomly selected 2D circular regions of the image volume are selected, a third annotator marks all vessels visible in the 2D region \"
\nNow it is a bit late. It might be better if these circular region annotation are released as well at the start of the competition.
\nE.g. these can help to verfiy a pseudo label system.
lesson learn or future plan?
\nsince all machine learning system mimick human action, it may be useful to relase results of different annotation steps. These actually provide additinal infomation (e.g. difficulty of label, etc)
\nan example is \"chain of thought\" and \"human feedback\" in LLM
\nby the way, is kidney1 voi label one of the \"circular region annotation\"?
", "rawMarkdown": "yashvrdnjain Thanks! I missed that early post.\n\n---\n\n```\n\"Hi so, here are some answers that will hopefully help:\n\nYes\nBM05 is a beamline at the synchrotron where the data are collected, (nothing to do with segmentation).\nYes it is manually segmented by annotators. We used a two annotator and validation strategy. After the first annotation by an experienced annotator is done, a second, independent, experienced annotator meticulously conducts 3D proofreading of the binary labels filling, in any missing vessel labels in the three orthogonal planes. After this, five randomly selected 2D circular regions of the image volume are selected, a third annotator marks all vessels visible in the 2D region (1 mark per vessel) then, using the labels, counts percentage of vessels that were missed i.e. the FN. This gives our % segmentation that we provide. For the complete segmentations, we just keep going around this loop iteratively to capture all vessels.\"\n```\n---\n\n\" five randomly selected 2D circular regions of the image volume are selected, a third annotator marks all vessels visible in the 2D region \"\n\nNow it is a bit late. It might be better if these circular region annotation are released as well at the start of the competition.\nE.g. these can help to verfiy a pseudo label system.\n\n---\n\nlesson learn or future plan?\n\nsince all machine learning system mimick human action, it may be useful to relase results of different annotation steps. These actually provide additinal infomation (e.g. difficulty of label, etc)\n\nan example is \"chain of thought\" and \"human feedback\" in LLM\n\n---\n\nby the way, is kidney1 voi label one of the \"circular region annotation\"?", "votes": null } ], "comments": [ { "id": 2587146, "author_name": "igorkrashenyi", "author_url": "", "post_date": "01/04/2024 15:22:15", "content": "This means that just 65% of labels are present in the ground truth images.
", "votes": null, "replies": [ { "id": 2587357, "author_name": "hengck23", "author_url": "", "post_date": "01/04/2024 17:56:21", "content": "i wonder how did the annotator knows 65% are present?
\nhe must know all 100 % first to compute the percentage of label present.
\nin that case, why don't he just let us have the 100%?
", "votes": null, "replies": [ { "id": 2587380, "author_name": "yashvrdnjain", "author_url": "", "post_date": "01/04/2024 18:05:59", "content": "Please see: https://www.kaggle.com/competitions/blood-vessel-segmentation/discussion/454353
\nHope this helps!
@yashvrdnjain Thanks! I missed that early post.
\n\n\" five randomly selected 2D circular regions of the image volume are selected, a third annotator marks all vessels visible in the 2D region \"
\nNow it is a bit late. It might be better if these circular region annotation are released as well at the start of the competition.
\nE.g. these can help to verfiy a pseudo label system.
lesson learn or future plan?
\nsince all machine learning system mimick human action, it may be useful to relase results of different annotation steps. These actually provide additinal infomation (e.g. difficulty of label, etc)
\nan example is \"chain of thought\" and \"human feedback\" in LLM
\nby the way, is kidney1 voi label one of the \"circular region annotation\"?
", "votes": null, "replies": [] } ] } ] } ] } ], "raw_markdown_by_id": { "2586727": "What does 'Sparsely segmented (about 65%)' mean?\n\n>kidney_2 - The whole of a kidney from another donor, at 50um resolution. Sparsely segmented (about 65%).\n\n", "2587146": "This means that just 65% of labels are present in the ground truth images.", "2587357": "i wonder how did the annotator knows 65% are present?\n\nhe must know all 100 % first to compute the percentage of label present.\n\nin that case, why don't he just let us have the 100%?", "2587380": "Please see: https://www.kaggle.com/competitions/blood-vessel-segmentation/discussion/454353 \nHope this helps!", "2587548": "yashvrdnjain Thanks! I missed that early post.\n\n---\n\n```\n\"Hi so, here are some answers that will hopefully help:\n\nYes\nBM05 is a beamline at the synchrotron where the data are collected, (nothing to do with segmentation).\nYes it is manually segmented by annotators. We used a two annotator and validation strategy. After the first annotation by an experienced annotator is done, a second, independent, experienced annotator meticulously conducts 3D proofreading of the binary labels filling, in any missing vessel labels in the three orthogonal planes. After this, five randomly selected 2D circular regions of the image volume are selected, a third annotator marks all vessels visible in the 2D region (1 mark per vessel) then, using the labels, counts percentage of vessels that were missed i.e. the FN. This gives our % segmentation that we provide. For the complete segmentations, we just keep going around this loop iteratively to capture all vessels.\"\n```\n---\n\n\" five randomly selected 2D circular regions of the image volume are selected, a third annotator marks all vessels visible in the 2D region \"\n\nNow it is a bit late. It might be better if these circular region annotation are released as well at the start of the competition.\nE.g. these can help to verfiy a pseudo label system.\n\n---\n\nlesson learn or future plan?\n\nsince all machine learning system mimick human action, it may be useful to relase results of different annotation steps. These actually provide additinal infomation (e.g. difficulty of label, etc)\n\nan example is \"chain of thought\" and \"human feedback\" in LLM\n\n---\n\nby the way, is kidney1 voi label one of the \"circular region annotation\"?" }, "source": "meta" }