rodriguescarson/eligible-scroll-atlas-renders

Get one mesh in about twenty seconds

curl -sO https://raw.githubusercontent.com/rodriguescarson/eligible-scroll-atlas/main/scripts/atlas.py
python atlas.py list --ink-pass                      # the 5 meshes that pass the pre-registered screen
python atlas.py ink PHerc0125 z10544_w020 --preview  # a downsampled ink map, about 12 KB
python atlas.py get PHerc0125 z10544_w020            # the surface volume, 31 planes, plane 15 is the surface

from atlas import meshes, fetch_surface_volume, open_surface
row = [m for m in meshes() if m["ink_pass_v1"] == "True"][0]
volume = open_surface(fetch_surface_volume(row["scroll"], row["mesh"], "./atlas"))
volume.shape        # (31, H, W) uint8

manifest.csv in this repository has one row per mesh: area, the eye verdict, sheet alignment, both models' scores, the Hecate rank, and a direct URL for the mesh's tar and ink-detection folder. The maps_held column marks the five meshes whose maps are held back, described below.

Eligible scroll atlas: surface-volume renders and raw ink_9um maps for the 340 published meshes of the eight prize-eligible 9 µm scrolls

Renders of every mesh in pscamillo/vesuvius-eligible-meshes (commit 0c966f8f) into the Scroll Prize team's segment layout, plus the raw output of three public ink_9um checkpoint files in both layer directions. Pre-registered recipe, screens and outcomes: rodriguescarson/eligible-scroll-atlas (prereg/PREREG.md, tag prereg-v1; prereg/AMENDMENT-1.md, tag prereg-v2).

Layout, one folder per scroll and mesh (the zarr is shipped as one tar per mesh because the raw layout is 580k chunk files, past the Hub's per-repo limit):

<scroll>/<mesh>/surface-volumes.tar     untar inside <scroll>/<mesh>/ to get:
    surface-volumes/<um>um-1.2m-<keV>keV-volume-<volid>.zarr   31 layers, uint8, chunks 31x128x128, zstd, 6-level pyramid
    render.json                                                 command, AppImage sha256, villa and mesh commits, wall time
<scroll>/<mesh>/ink-detection/<scroll>-<mesh>-<volid>-ink9um-<tag>.tif      raw uint8 probability map, forward layer order
<scroll>/<mesh>/ink-detection/...-<tag>_reverse.tif                          same, reversed layer order
<scroll>/<mesh>/ink-detection/*-ds8.jpg                                      8x block mean, displayed as (p-0.25)/0.5
<scroll>/<mesh>/ink-detection/valid_mask.tif                                 nonzero at level 0, eroded 64 px
<scroll>/<mesh>/ink-detection/infer.json                                     recipe, checkpoint tags, villa commit

Render: vc_render_tifxyz from VC3D-4b3c728-2026-09-15-linux-x86_64.AppImage, --scale 1 --group-idx 0 --num-slices 31 --slice-step 1 --flip-normals, voxel size 9.362 µm (8.64 for PHerc0800 and PHerc0268), each scroll's published masked volume by URL. --flip-normals is the setting that reproduces the team's own PHerc0800 surface volume (Pearson r 1.0000, repro/ in the code repo). Inference: vesuvius.ink_detection.inference.infer at villa 4b3c728, --overlap 0.5 --blend-mode hann --direction both, checkpoints hybrid_3d2d-seed43/step-060000 (s43_060k), hybrid_3d2d-seed42/step-020000 (s42_020k) and a uniform weight average of seed42 steps 10000/20000/30000 (soup42_early3, recipe in scripts/soup42.py).

These are raw model outputs, not detections. Nothing here has been claimed as text.

Licence: the renders and maps are derived from Vesuvius Challenge open data, which is released under CC BY-NC 4.0; these files carry the same terms. The code in the GitHub repository is MIT.

Maps are published scroll by scroll after the pre-registered screens have run on them; a mesh that passes the screens is listed with its scores only and its images go to the Scroll Prize team first (PREREG, "Outcomes").

The known-ink control (PHerc0139 w043) and the two planted-window controls are under controls/.