SharpMeow's picture
Refresh: 2290 meshes of 329 segments, 0 not measured (kit a846745, 9 s)
c0db88a verified
|
Raw History Blame Contribute Delete
4.21 kB
---
license: cc-by-nc-4.0
pretty_name: Vesuvius Segment Atlas
tags:
- vesuvius-challenge
- herculaneum
- tifxyz
- segmentation
configs:
- config_name: default
data_files:
- split: train
path: atlas.csv
---
# Vesuvius Segment Atlas
The same geometry numbers for every traced segment the [Vesuvius Challenge open data](https://vesuvius-challenge-open-data.s3.us-east-1.amazonaws.com/) publishes:
one row per tifxyz mesh (a segment mapped to one scan volume), with its area in cm2, how much of its grid is filled,
how isometric the grid is, and how many grid nodes sit in holes inside the sheet, in folds, or on over-stretched
edges. Built by `python -m kit atlas build` from [ChaseHendrick/Scrolls](https://github.com/ChaseHendrick/Scrolls)
(page: `docs/tools/atlas.md`), refreshed daily by a Hugging Face Job that measures only new or changed meshes.
Built 2026-11-19T06:17:59Z from commit `a846745`: 2290 meshes of 329
segments, 0 not measured (the `error` column says why).
| Scroll | Segments | Meshes | Measured | Clean | Area, cm2 (each segment once) | Prize lists (kit snapshot) |
| --- | --- | --- | --- | --- | --- | --- |
| PHerc0009B | 18 | 90 | 90 | 42 | 350.3 | none |
| PHerc0139 | 39 | 607 | 607 | 31 | 1207.5 | none |
| PHerc0172 | 53 | 212 | 212 | 0 | 7067.3 | none |
| PHerc0211 | 1 | 3 | 3 | 0 | 7.7 | first-letters-2027, grand-prize-2027 |
| PHerc0306B | 1 | 3 | 3 | 0 | 23.9 | none |
| PHerc0332 | 2 | 2 | 2 | 0 | 0.0 | none |
| PHerc0343 | 1 | 3 | 3 | 0 | 7.6 | none |
| PHerc0343P | 8 | 40 | 40 | 32 | 32.7 | none |
| PHerc0500P2 | 46 | 342 | 342 | 134 | 262.7 | none |
| PHerc0800 | 6 | 18 | 18 | 15 | 9.5 | first-letters-2027, grand-prize-2027 |
| PHerc0814 | 19 | 125 | 125 | 6 | 470.4 | none |
| PHerc0841 | 3 | 15 | 15 | 0 | 36.4 | none |
| PHerc1447 | 18 | 54 | 54 | 0 | 93.9 | grand-prize-2027 |
| PHerc1667 | 20 | 159 | 159 | 0 | 1860.9 | none |
| PHercMAN5 | 1 | 1 | 1 | 0 | 0.0 | none |
| PHercMANBp | 11 | 11 | 11 | 0 | 0.0 | none |
| PHercParis4 | 82 | 605 | 605 | 0 | 45096.3 | first-title |
## What the numbers mean
- **area_cm2**: two triangles per complete grid cell, in voxels squared, times the voxel size squared. The voxel
size comes from the folder name (`-N.NNNum.tifxyz`), else from the volume's metadata, else from the scroll's only
volume (`voxel_um_from` says which); without one the area is empty. **meta_area_ratio** is our area over the
`area_vx2` VC3D wrote into `meta.json`, where there is one.
- **edge_p5, edge_p50, edge_p95**: every grid edge's 3D length over the step `meta.json`'s `scale` promises (1.0 is
exact); **within_10pct** and **within_25pct** are the shares of edges that close; **median_abs_log** is the median
of the absolute log of that ratio; **cell_area_p50** is cell area over the promised area; **angle_p5/p50/p95** is
the angle between the two grid directions, in degrees (90 is square).
- **holes_nodes**: missing nodes not joined to the grid's border through missing nodes (holes inside the sheet).
**folds_nodes**: nodes of neighboring cells whose normals turn by more than 90 degrees. **stretched_nodes**: nodes
on an edge more than 1.5 times too long or too short. Each also in mm2. **clean** is true when all three are 0.
- **kind** is `mesh` for a top-level mesh folder and the subfolder name (such as `intermediate`) otherwise. Area per
scroll in the table counts each segment once, from its finest-voxel top-level mesh.
- **lens**: for scrolls on no prize list, the file in `lens/` with a small preview grid used by the Segment Lens
viewer (format in `docs/tools/atlas.md`). Scrolls on a prize list get numbers only.
- **fingerprint** comes from the ETags and sizes of x.tif, y.tif and z.tif; a refresh measures a mesh again only when
it changes.
These numbers describe the grid, not the papyrus. A clean mesh can still jump from one sheet to the next; nothing
here looks at the CT or judges ink, and none of it is an official villa or prize check.
## License and credit
The meshes are Vesuvius Challenge open data, CC BY-NC 4.0; these derived numbers and preview grids carry the same
license. Credit: Vesuvius Challenge open data, CC BY-NC 4.0. The code that makes them is Apache-2.0.