Snapped Labels

What a snap does

A surface label is a mesh (tifxyz: three float grids, x, y and z in scan voxels) that follows one wrap of papyrus. Traced labels sit near the sheet but rarely on it: on fresh test sites only about a quarter of published label vertices were within one voxel of the sheet's recto face. A snap moves each vertex along its mesh normal onto the face of the bright band nearest it, on the side toward the scroll's center line.

Two methods

• finesnap (PHerc0139, PHerc0814, PHerc0841): the label's own 9.4 µm scan is registered, vertex by vertex, to a 2.4 µm scan of the same scroll, and the face is found there. Where that fails, the label scan's own face is used (source.tif says which).
• labelsnap (PHerc0172): the face is found in the label's own 7.9 µm scan.

How it was measured

Truth is the recto face seen in an independent 1.129 µm scan, at test sites drawn after the method was fixed and never used before. PHerc0139, 60 sites: within 1 voxel 23.7% for the published labels, 79.4% for labelsnap, 91.2% for finesnap. PHerc0814, 60 sites: 24.2% for published labels, 71.0% for labelsnap. A 3D surface model trained on snapped PHerc0139 labels placed the surface 0.61 voxels closer on PHerc0814 than the same model trained on published labels (95% interval 0.39 to 0.84; all 9 seed pairings).

Against seven community snaps

On 29 October 2026 the single-scan snap (labelsnap, the method used here for PHerc0172) ran once on fresh test sites of PHerc0814 and PHerc0841 next to seven community snaps, under a rule fixed before it ran. Truth is the recto face in a finer scan of the same scroll (1.129 µm for PHerc0814, 2.403 µm for PHerc0841), registered per site; 71 sites were scored. Pooled, labelsnap moved labels 2.15 voxels closer to that face than the published labels (95% interval 1.82 to 2.51) and 1.89 closer than VC3D's vc_objrefine with thresholds tuned on PHerc0139 (1.48 to 2.35). It was also closer than vc_objrefine as shipped, recto-offset r2 with and without its gate, surface_placement_audit, RectoRefine's midline and a snap onto a surface prediction, on each scroll and pooled. Within 1 voxel: 55% (PHerc0814) and 66% (PHerc0841) for labelsnap, 19% for the published labels on both, 22% and 21% for the best community snap. Every control held: reversing the move pushed labels away, a random sign gained nothing.

Limits: PHerc0814 is held out by site, not by scroll. The test shows labels moved closer to the face, not every vertex within 1 voxel. A label displaced toward the verso was not shown to come back. The midline and nearest-peak snaps aim at the sheet center by design, so part of their loss is that target. Where a finer scan exists, finesnap did better than labelsnap on PHerc0814 by 0.93 voxels. The released PHerc0841 meshes were not scored here, because their finesnap read the same 2.403 µm scan this test used as truth.

What is in each segment folder

• snapped.tifxyz/: the moved mesh, same grid as the source; -1 where there is no vertex or it was held back.
• abstain.tif: 0 snapped (face measured at the vertex), 1 not measured and left at its published position, 2 not measured and moved by the offset filled in from measured neighbors (filled), 255 no vertex or held back. Version 1.0.1 added code 2; in 1.0.0 filled vertices were 1.
• confidence.tif (0 to 1), reason.tif (why no face was measured), source.tif (finesnap: 1 fine scan, 2 label scan).
• release_mask.tif: 0 released, 1 held back for evaluation, 255 not in the source mesh. release.json: source URLs, checksums and the summary.

Limits

A snap moves along the normal only; it does not fix a label that follows the wrong wrap. Where two sheets touch, the face it finds can belong to the neighbor. Small regions are held back for evaluation and five segments of PHerc0139 and one of PHerc0841 are not included. The meshes are geometry: they say nothing about ink or text.

Scans and source meshes: Vesuvius Challenge open data, CC BY-NC 4.0; the snapped meshes are shared under the same terms. Page code Apache 2.0. Made by Chase Hendrick.