{ "id": 459844, "title": "Sparse & Dense segmentations? Clarification 😅", "url": "/competitions/blood-vessel-segmentation/discussion/459844", "author_name": "", "post_date": "2023-12-06T19:01:09.539148Z", "votes": 4, "comment_count": 6, "views": 0, "content": "

I'm a noob in segmentation problems and don't really understand a lot of the vocabulary being used here. I just wanted to ask what exactly is the purpose of dense and sparse segmentations that have been provided?
\nIt might sound trivial but yep 😅

", "messages": [ { "id": "2551549", "postDate": "12/06/2023 19:01:09", "content": "

I'm a noob in segmentation problems and don't really understand a lot of the vocabulary being used here. I just wanted to ask what exactly is the purpose of dense and sparse segmentations that have been provided?
\nIt might sound trivial but yep 😅

", "rawMarkdown": "I'm a noob in segmentation problems and don't really understand a lot of the vocabulary being used here. I just wanted to ask what exactly is the purpose of dense and sparse segmentations that have been provided?\nIt might sound trivial but yep 😅", "votes": null }, { "id": "2551572", "postDate": "12/06/2023 19:25:18", "content": "

You can think that every segmented object in the dataset as a 3D grid or like a big cube made of little cubes. Each little cube is called a voxel, the 3-dimensional counterpart of a pixel (2-dimensional). The voxels come in a certain resolution. Here, in this competition, the resolution is 50 um, that means that each voxel's side measures 50 micrometers. Each voxel can have 1s or 0s. If the object you want to segment is present in that portion of space it has a 1, otherwise a 0.

\n

That being said, you could segment the entire volume of the object (100%), that is, a dense segmentation. In contrast, I believe that a sparse segmentation is a segmentation where not all parts of the 3D object is segmented but rather parts of it, a subset of voxels of an object, typically those that correspond to the boundaries or key objects of interest.

\n

Why do we have sparse segmentations? I do not entirely know, maybe it was because of the measurement device used or memory constraints.

\n

I may be wrong, but that is how I understand it.

", "rawMarkdown": "You can think that every segmented object in the dataset as a 3D grid or like a big cube made of little cubes. Each little cube is called a **voxel**, the 3-dimensional counterpart of a **pixel** (2-dimensional). The voxels come in a certain resolution. Here, in this competition, the resolution is 50 um, that means that each voxel's side measures 50 micrometers. Each voxel can have 1s or 0s. If the object you want to segment is present in that portion of space it has a 1, otherwise a 0.\n\nThat being said, you could segment the entire volume of the object (100%), that is, a **dense** segmentation. In contrast, I believe that a **sparse** segmentation is a segmentation where not all parts of the 3D object is segmented but rather parts of it, a subset of voxels of an object, typically those that correspond to the boundaries or key objects of interest.\n\nWhy do we have sparse segmentations? I do not entirely know, maybe it was because of the measurement device used or memory constraints.\n\nI may be wrong, but that is how I understand it.", "votes": null }, { "id": "2551578", "postDate": "12/06/2023 19:30:32", "content": "

Thanks for the elaborate description
\nI finally understood the relevance of resolution

\n

But still I have a doubt
\nIf sparse segmentation refers to a \"part\" of the whole, then why do we have ~500 labels in dense folder and more than 1000 in sparse folder😬. Shouldn't dense labels be more in number?

", "rawMarkdown": "Thanks for the elaborate description\nI finally understood the relevance of resolution\n\nBut still I have a doubt\nIf sparse segmentation refers to a \"part\" of the whole, then why do we have ~500 labels in dense folder and more than 1000 in sparse folder😬. Shouldn't dense labels be more in number?", "votes": null }, { "id": "2551585", "postDate": "12/06/2023 19:39:42", "content": "

I have not entered this competition so I am not familiar with the data. But to illustrate sparse and dense segmentations this image might be useful:

\n

\n

My undestanding is that the upper car is densely segmented and the lower car is sparsely segmented. Both objects would have the same dimensions, but the sparsely segmented one has less 1s compared to the densely segmented one.

", "rawMarkdown": "I have not entered this competition so I am not familiar with the data. But to illustrate sparse and dense segmentations this image might be useful:\n\n\n\nMy undestanding is that the upper car is densely segmented and the lower car is sparsely segmented. Both objects would have the same dimensions, but the sparsely segmented one has less 1s compared to the densely segmented one.", "votes": null }, { "id": "2553383", "postDate": "12/08/2023 08:09:41", "content": "

I was hoping to see some explainatioj done in the comments
\nAbout why are there lesser labels in dense fllder compared to in the sparse folder

", "rawMarkdown": "I was hoping to see some explainatioj done in the comments\nAbout why are there lesser labels in dense fllder compared to in the sparse folder", "votes": null }, { "id": "2553507", "postDate": "12/08/2023 10:05:16", "content": "

This is specific to this competition. From what I understood, as it's very hard to perfectly segment a 3D stack, organizers implemented an annotation/review process : after annotation, another reviewer looked at some random slices and pointed the missing labeled vessels and correct vessels -> this creates a score of annotation with the ratio of segmented vessels. They iterated through annotation and review to make the ratio go up to 100%. Some stacks reached 100% (dense), some 85% (sparse) and some 65% (sparser).
\nRead the preprint for more information.

", "rawMarkdown": "This is specific to this competition. From what I understood, as it's very hard to perfectly segment a 3D stack, organizers implemented an annotation/review process : after annotation, another reviewer looked at some random slices and pointed the missing labeled vessels and correct vessels -> this creates a score of annotation with the ratio of segmented vessels. They iterated through annotation and review to make the ratio go up to 100%. Some stacks reached 100% (dense), some 85% (sparse) and some 65% (sparser).\nRead the [preprint](https://www.kaggle.com/competitions/blood-vessel-segmentation/discussion/456761) for more information.", "votes": null }, { "id": "2554162", "postDate": "12/08/2023 20:50:06", "content": "

Thanks
\nThis makes so much more sense now

", "rawMarkdown": "Thanks\nThis makes so much more sense now", "votes": null } ], "comments": [ { "id": 2551572, "author_name": "alejopaullier", "author_url": "", "post_date": "12/06/2023 19:25:18", "content": "

You can think that every segmented object in the dataset as a 3D grid or like a big cube made of little cubes. Each little cube is called a voxel, the 3-dimensional counterpart of a pixel (2-dimensional). The voxels come in a certain resolution. Here, in this competition, the resolution is 50 um, that means that each voxel's side measures 50 micrometers. Each voxel can have 1s or 0s. If the object you want to segment is present in that portion of space it has a 1, otherwise a 0.

\n

That being said, you could segment the entire volume of the object (100%), that is, a dense segmentation. In contrast, I believe that a sparse segmentation is a segmentation where not all parts of the 3D object is segmented but rather parts of it, a subset of voxels of an object, typically those that correspond to the boundaries or key objects of interest.

\n

Why do we have sparse segmentations? I do not entirely know, maybe it was because of the measurement device used or memory constraints.

\n

I may be wrong, but that is how I understand it.

", "votes": null, "replies": [ { "id": 2551578, "author_name": "bhavesjain", "author_url": "", "post_date": "12/06/2023 19:30:32", "content": "

Thanks for the elaborate description
\nI finally understood the relevance of resolution

\n

But still I have a doubt
\nIf sparse segmentation refers to a \"part\" of the whole, then why do we have ~500 labels in dense folder and more than 1000 in sparse folder😬. Shouldn't dense labels be more in number?

", "votes": null, "replies": [ { "id": 2551585, "author_name": "alejopaullier", "author_url": "", "post_date": "12/06/2023 19:39:42", "content": "

I have not entered this competition so I am not familiar with the data. But to illustrate sparse and dense segmentations this image might be useful:

\n

\n

My undestanding is that the upper car is densely segmented and the lower car is sparsely segmented. Both objects would have the same dimensions, but the sparsely segmented one has less 1s compared to the densely segmented one.

", "votes": null, "replies": [] } ] } ] }, { "id": 2553383, "author_name": "gudia3110", "author_url": "", "post_date": "12/08/2023 08:09:41", "content": "

I was hoping to see some explainatioj done in the comments
\nAbout why are there lesser labels in dense fllder compared to in the sparse folder

", "votes": null, "replies": [] }, { "id": 2553507, "author_name": "optimo", "author_url": "", "post_date": "12/08/2023 10:05:16", "content": "

This is specific to this competition. From what I understood, as it's very hard to perfectly segment a 3D stack, organizers implemented an annotation/review process : after annotation, another reviewer looked at some random slices and pointed the missing labeled vessels and correct vessels -> this creates a score of annotation with the ratio of segmented vessels. They iterated through annotation and review to make the ratio go up to 100%. Some stacks reached 100% (dense), some 85% (sparse) and some 65% (sparser).
\nRead the preprint for more information.

", "votes": null, "replies": [ { "id": 2554162, "author_name": "bhavesjain", "author_url": "", "post_date": "12/08/2023 20:50:06", "content": "

Thanks
\nThis makes so much more sense now

", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "2551549": "I'm a noob in segmentation problems and don't really understand a lot of the vocabulary being used here. I just wanted to ask what exactly is the purpose of dense and sparse segmentations that have been provided?\nIt might sound trivial but yep 😅", "2551572": "You can think that every segmented object in the dataset as a 3D grid or like a big cube made of little cubes. Each little cube is called a **voxel**, the 3-dimensional counterpart of a **pixel** (2-dimensional). The voxels come in a certain resolution. Here, in this competition, the resolution is 50 um, that means that each voxel's side measures 50 micrometers. Each voxel can have 1s or 0s. If the object you want to segment is present in that portion of space it has a 1, otherwise a 0.\n\nThat being said, you could segment the entire volume of the object (100%), that is, a **dense** segmentation. In contrast, I believe that a **sparse** segmentation is a segmentation where not all parts of the 3D object is segmented but rather parts of it, a subset of voxels of an object, typically those that correspond to the boundaries or key objects of interest.\n\nWhy do we have sparse segmentations? I do not entirely know, maybe it was because of the measurement device used or memory constraints.\n\nI may be wrong, but that is how I understand it.", "2551578": "Thanks for the elaborate description\nI finally understood the relevance of resolution\n\nBut still I have a doubt\nIf sparse segmentation refers to a \"part\" of the whole, then why do we have ~500 labels in dense folder and more than 1000 in sparse folder😬. Shouldn't dense labels be more in number?", "2551585": "I have not entered this competition so I am not familiar with the data. But to illustrate sparse and dense segmentations this image might be useful:\n\n\n\nMy undestanding is that the upper car is densely segmented and the lower car is sparsely segmented. Both objects would have the same dimensions, but the sparsely segmented one has less 1s compared to the densely segmented one.", "2553383": "I was hoping to see some explainatioj done in the comments\nAbout why are there lesser labels in dense fllder compared to in the sparse folder", "2553507": "This is specific to this competition. From what I understood, as it's very hard to perfectly segment a 3D stack, organizers implemented an annotation/review process : after annotation, another reviewer looked at some random slices and pointed the missing labeled vessels and correct vessels -> this creates a score of annotation with the ratio of segmented vessels. They iterated through annotation and review to make the ratio go up to 100%. Some stacks reached 100% (dense), some 85% (sparse) and some 65% (sparser).\nRead the [preprint](https://www.kaggle.com/competitions/blood-vessel-segmentation/discussion/456761) for more information.", "2554162": "Thanks\nThis makes so much more sense now" }, "source": "meta" }