Hello all,\nIs there any information about the creation of the dataset. \nI know they are crops from much larger volumes. I would like to know if we have access to the entire volumes and how was the crop selection done as well as the segmentation process.\nMaybe the data comes from here: https://scrollprize.org/unwrapping\nThank you for your time!
\nBest regards,\nAndré Ferreira
", "messages": [ { "id": "3349302", "postDate": "11/26/2025 15:33:44", "content": "Hello all,\nIs there any information about the creation of the dataset. \nI know they are crops from much larger volumes. I would like to know if we have access to the entire volumes and how was the crop selection done as well as the segmentation process.\nMaybe the data comes from here: https://scrollprize.org/unwrapping\nThank you for your time!
\nBest regards,\nAndré Ferreira
", "rawMarkdown": "Hello all,\nIs there any information about the creation of the dataset. \nI know they are crops from much larger volumes. I would like to know if we have access to the entire volumes and how was the crop selection done as well as the segmentation process.\nMaybe the data comes from here: https://scrollprize.org/unwrapping\nThank you for your time!\n\nBest regards,\nAndré Ferreira", "votes": null }, { "id": "3349594", "postDate": "11/26/2025 18:25:04", "content": "I found a case where the segmentation seems to be incorrect, although I'm not really sure if the problem was interpolation. File with name 1407735.tif, slice 160.
", "rawMarkdown": "I found a case where the segmentation seems to be incorrect, although I'm not really sure if the problem was interpolation. File with name 1407735.tif, slice 160.", "votes": null }, { "id": "3349705", "postDate": "11/26/2025 22:50:46", "content": "Hello! ,
\nThe dataset is created by a couple methods, but all are derived from 'segmentations'. I put segmentations in quotes here as we use the term a bit weird. A segmentation in this context is a single mesh from a portion of the scroll which was created either manually (using khartes or legacy volume-cartographer) or semi/fully automatically using vc3d.
\nThe recto portion of the scroll is traced and meshed and then parameterized into a 2d surface using slim or abf. To generate labels, these meshes are voxelized (by rasterizing the 3d triangles and connecting disconnected pieces with a simple line algorithm). From there, I went through each zarr one chunk of size 320^3 at a time and identified areas which had \"clean\" labels, ideally grouped together.
\nThe volumes in the training set come from a combination of public and yet to be released data. We currently host data in two places:
\nash2txt file server:\nin this directory you will find directories for five scrolls, and in the volumes/ directory within them are either zarrs or single 2d tif slices
\nfragments:\nthis directory contains fragments (pieces of unrolling attempts)
\nwe also have two more volumes here , which will be where we our moving our data in the future.
\nthe labeled data represents and infinitely small fraction of the total volumes
\nthe segmentation guide on our website should give you a decent introduction into segmentation with vc3d, at least enough to start producing these. if you wish to get into segmentation and run into any issues don't hesitate to ask
", "rawMarkdown": "Hello! ,\n\nThe dataset is created by a couple methods, but all are derived from 'segmentations'. I put segmentations in quotes here as we use the term a bit weird. A segmentation in this context is a single mesh from a portion of the scroll which was created either manually (using khartes or legacy volume-cartographer) or semi/fully automatically using vc3d. \n\nThe recto portion of the scroll is traced and meshed and then parameterized into a 2d surface using slim or abf. To generate labels, these meshes are voxelized (by rasterizing the 3d triangles and connecting disconnected pieces with a simple line algorithm). From there, I went through each zarr one chunk of size 320^3 at a time and identified areas which had \"clean\" labels, ideally grouped together. \n\nThe volumes in the training set come from a combination of public and yet to be released data. We currently host data in two places: \n\n[ash2txt file server](https://dl.ash2txt.org/full-scrolls):\nin this directory you will find directories for five scrolls, and in the volumes/ directory within them are either zarrs or single 2d tif slices\n\n[fragments](https://dl.ash2txt.org/fragments/):\nthis directory contains fragments (pieces of unrolling attempts)\n\nwe also have two more volumes [here](https://data.aws.ash2txt.org/samples/) , which will be where we our moving our data in the future. \n\nthe labeled data represents and infinitely small fraction of the total volumes\n\nthe segmentation guide on [our website](www.scrollprize.org/segmentation) should give you a decent introduction into segmentation with vc3d, at least enough to start producing these. if you wish to get into segmentation and run into any issues don't hesitate to ask", "votes": null }, { "id": "3350366", "postDate": "11/27/2025 14:17:08", "content": "Thank you for the detailed explanation! It's definitly a lot of work to create the labels, thanks!
\nJust one more thing to make sure I have the correct mental image of the process. When we load a standard chunk (e.g., $320^3$ voxels), is this effectively a direct, 1:1 crop from the original high-resolution reconstruction? Meaning: If the native resolution is ~8µm, this chunk represents a tiny physical cube of roughly 2.5mm x 2.5mm x 2.5mm cut from the scroll, without any resizing or downsampling? I.e. basically zooming in and crop? I just want to ensure these aren't downscaled overviews, but rather full-resolution 'micro-crops'.
", "rawMarkdown": "Thank you for the detailed explanation! It's definitly a lot of work to create the labels, thanks!\n\nJust one more thing to make sure I have the correct mental image of the process. When we load a standard chunk (e.g., $320^3$ voxels), is this effectively a direct, 1:1 crop from the original high-resolution reconstruction? Meaning: If the native resolution is ~8µm, this chunk represents a tiny physical cube of roughly 2.5mm x 2.5mm x 2.5mm cut from the scroll, without any resizing or downsampling? I.e. basically zooming in and crop? I just want to ensure these aren't downscaled overviews, but rather full-resolution 'micro-crops'.", "votes": null }, { "id": "3350801", "postDate": "11/27/2025 20:00:19", "content": "Correct, these have not been resampled at all , so the physical dimensions are accurate for the scroll they come from. They are chunks directly from the raw volumes.
", "rawMarkdown": "Correct, these have not been resampled at all , so the physical dimensions are accurate for the scroll they come from. They are chunks directly from the raw volumes.", "votes": null } ], "comments": [ { "id": 3349594, "author_name": "andrefilipeferreira", "author_url": "", "post_date": "11/26/2025 18:25:04", "content": "I found a case where the segmentation seems to be incorrect, although I'm not really sure if the problem was interpolation. File with name 1407735.tif, slice 160.
", "votes": null, "replies": [] }, { "id": 3349705, "author_name": "seanjohnsonsp", "author_url": "", "post_date": "11/26/2025 22:50:46", "content": "Hello! ,
\nThe dataset is created by a couple methods, but all are derived from 'segmentations'. I put segmentations in quotes here as we use the term a bit weird. A segmentation in this context is a single mesh from a portion of the scroll which was created either manually (using khartes or legacy volume-cartographer) or semi/fully automatically using vc3d.
\nThe recto portion of the scroll is traced and meshed and then parameterized into a 2d surface using slim or abf. To generate labels, these meshes are voxelized (by rasterizing the 3d triangles and connecting disconnected pieces with a simple line algorithm). From there, I went through each zarr one chunk of size 320^3 at a time and identified areas which had \"clean\" labels, ideally grouped together.
\nThe volumes in the training set come from a combination of public and yet to be released data. We currently host data in two places:
\nash2txt file server:\nin this directory you will find directories for five scrolls, and in the volumes/ directory within them are either zarrs or single 2d tif slices
\nfragments:\nthis directory contains fragments (pieces of unrolling attempts)
\nwe also have two more volumes here , which will be where we our moving our data in the future.
\nthe labeled data represents and infinitely small fraction of the total volumes
\nthe segmentation guide on our website should give you a decent introduction into segmentation with vc3d, at least enough to start producing these. if you wish to get into segmentation and run into any issues don't hesitate to ask
", "votes": null, "replies": [ { "id": 3350366, "author_name": "andrefilipeferreira", "author_url": "", "post_date": "11/27/2025 14:17:08", "content": "Thank you for the detailed explanation! It's definitly a lot of work to create the labels, thanks!
\nJust one more thing to make sure I have the correct mental image of the process. When we load a standard chunk (e.g., $320^3$ voxels), is this effectively a direct, 1:1 crop from the original high-resolution reconstruction? Meaning: If the native resolution is ~8µm, this chunk represents a tiny physical cube of roughly 2.5mm x 2.5mm x 2.5mm cut from the scroll, without any resizing or downsampling? I.e. basically zooming in and crop? I just want to ensure these aren't downscaled overviews, but rather full-resolution 'micro-crops'.
", "votes": null, "replies": [ { "id": 3350801, "author_name": "seanjohnsonsp", "author_url": "", "post_date": "11/27/2025 20:00:19", "content": "Correct, these have not been resampled at all , so the physical dimensions are accurate for the scroll they come from. They are chunks directly from the raw volumes.
", "votes": null, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "3349302": "Hello all,\nIs there any information about the creation of the dataset. \nI know they are crops from much larger volumes. I would like to know if we have access to the entire volumes and how was the crop selection done as well as the segmentation process.\nMaybe the data comes from here: https://scrollprize.org/unwrapping\nThank you for your time!\n\nBest regards,\nAndré Ferreira", "3349594": "I found a case where the segmentation seems to be incorrect, although I'm not really sure if the problem was interpolation. File with name 1407735.tif, slice 160.", "3349705": "Hello! ,\n\nThe dataset is created by a couple methods, but all are derived from 'segmentations'. I put segmentations in quotes here as we use the term a bit weird. A segmentation in this context is a single mesh from a portion of the scroll which was created either manually (using khartes or legacy volume-cartographer) or semi/fully automatically using vc3d. \n\nThe recto portion of the scroll is traced and meshed and then parameterized into a 2d surface using slim or abf. To generate labels, these meshes are voxelized (by rasterizing the 3d triangles and connecting disconnected pieces with a simple line algorithm). From there, I went through each zarr one chunk of size 320^3 at a time and identified areas which had \"clean\" labels, ideally grouped together. \n\nThe volumes in the training set come from a combination of public and yet to be released data. We currently host data in two places: \n\n[ash2txt file server](https://dl.ash2txt.org/full-scrolls):\nin this directory you will find directories for five scrolls, and in the volumes/ directory within them are either zarrs or single 2d tif slices\n\n[fragments](https://dl.ash2txt.org/fragments/):\nthis directory contains fragments (pieces of unrolling attempts)\n\nwe also have two more volumes [here](https://data.aws.ash2txt.org/samples/) , which will be where we our moving our data in the future. \n\nthe labeled data represents and infinitely small fraction of the total volumes\n\nthe segmentation guide on [our website](www.scrollprize.org/segmentation) should give you a decent introduction into segmentation with vc3d, at least enough to start producing these. if you wish to get into segmentation and run into any issues don't hesitate to ask", "3350366": "Thank you for the detailed explanation! It's definitly a lot of work to create the labels, thanks!\n\nJust one more thing to make sure I have the correct mental image of the process. When we load a standard chunk (e.g., $320^3$ voxels), is this effectively a direct, 1:1 crop from the original high-resolution reconstruction? Meaning: If the native resolution is ~8µm, this chunk represents a tiny physical cube of roughly 2.5mm x 2.5mm x 2.5mm cut from the scroll, without any resizing or downsampling? I.e. basically zooming in and crop? I just want to ensure these aren't downscaled overviews, but rather full-resolution 'micro-crops'.", "3350801": "Correct, these have not been resampled at all , so the physical dimensions are accurate for the scroll they come from. They are chunks directly from the raw volumes." }, "source": "meta" }