SharpMeow/vesuvius-center-lines

Center lines for the prize-eligible scans and more

Every public center line (umbilicus) we know of for the scans on the Vesuvius Challenge prize page (read 23 Oct 2026), copied unchanged into one folder: lines//_umbilicus.json, each in its scan's level-0 voxels (x, y, z), VC3D's umbilicus format.

index.json lists every file with its scan, source, license, upstream link, sha256, point count and z range. All 22 scans have at least one line.

Since v2.3 (24 Oct 2026) the dataset also covers 11 public scans that are not on the prize page, so 33 scans and 40 lines in all: the organizers' lines from the open-data bucket for PHerc0139, PHerc0332, PHerc1451 and PHercParis4 (copied unchanged), and new Center Finder lines for PHerc0343, PHerc0814 and PHerc0841 (1.2 m scans) and PHerc1299, PHercMAN5, PHercMANB and PHercParis3 (2.4 um scans, read in the finder's 2.4 um mode). These rows have first_letters and grand_prize false.

For the dataset viewer and SQL console, tables/lines.jsonl has one row per line (the index.json fields plus the CT end check grades) and tables/points.jsonl one row per control point (x, y, z). Both are built from index.json and the line files; the files under lines/ stay the source.

Sources: Alexey Drobkov (MIT), Iyan Dopico via Drobkov's repository (MIT), Abd Elilah's manual curves (CC BY 4.0), the Vesuvius Challenge open data (CC BY-NC 4.0), and lines from my own Center Finder for the nine 1.2 m scans nobody had done (CC BY-NC 4.0; model output, end status in each file's note). Thanks to everyone who made these.

End checks on my nine lines (automatic only, 23 Oct 2026): 14 of 18 ends pass and 4 are flagged (PHerc0175B high, PHerc0306B low, PHerc0490B low, PHerc0846A high). Each end is compared with the scroll center the CT itself shows there, estimated from the scan alone (sheet orientations in three slices, never the line), 0, 1 and 2 mm inside the end, with a 4 mm bar. A planted fake tilt is caught at all 18 ends. The check was tuned only on the 13 published human-made lines, where it catches the fake tilt at all 26 ends and passes 20 of them. 5 of the 14 passing ends were recapped: their last 10 mm replaced by a straight fit to the CT at heights the check doesn't judge (points marked ct-recap), so those passes say the CT supports that end, not that a separate check confirmed it. Model output: no end has been checked on CT by a person. usable_ends in each file's metadata says which ends to rely on; ct_end_check has the numbers.

End checks on the seven v2.3 Center Finder lines (automatic only, 24 Oct 2026, the same check and bar): all 14 ends pass, 2 of them after a CT recap (PHerc0343 low, PHercMANB high). The planted fake tilt was caught at all 14. The check was tuned only on 1.2 m scans, so on the four 2.4 um scans its grades are untested and no end is marked usable there (ct_untested_scan in index.json); on PHerc0343, PHerc0814 and PHerc0841 all six ends are usable. Model output: no end has been checked on CT by a person.

Since v2.4 (24 Oct 2026) there is also a polished copy of each of my 12 Center Finder lines on 1.2 m scans, beside the original (source finder-polished, file ..._finder-polished_umbilicus.json, polished_from names the original). The polish (kit axispolish, rule 1: https://github.com/ChaseHendrick/Scrolls/blob/main/docs/tools/axispolish.md) moves the line every 100 slices to the nearby point where villa's predicted sheet winds most around it. In one frozen test on ten fresh scans, the polished lines sat a little deeper in villa's traced sheet than villa's vc_gen_umbilicus (+0.084 wraps pooled, interval +0.028 to +0.169, 8 of 10 scans); the lines before the polish sat shallower (-0.330 wraps). That test measured coverage of villa's own model output, not distance to the true axis. Ten of the 12 scans were in it; PHerc0343 and PHerc0846A were not, so the polish is untested there (in_frozen_test). Two cautions. The polish moves each slice on its own, so these lines are jumpier than the originals: steps_over_4mm and max_step_mm in index.json count the sideways jumps between points (on PHerc0490A, 27 of 83 steps move over 4 mm, the largest 13.3 mm), and the test never scored smoothness. And the ends were not re-checked, so the CT end check of the original does not carry over. The original lines stay as they were. Where both exist, the polished line is the recommended default (Scrolls kit centerline get and AutoPipeline take it first); take the original (source finder) where a smooth axis or its CT-checked ends matter more, or smooth the polished line first. Model output: no person has checked a polished line on CT. Evidence: https://github.com/ChaseHendrick/Scrolls/blob/main/docs/evidence/2026-10-22-umbilicus-windpolish/README.md

Since v2.5 (24 Oct 2026) a line that has a newer version carries superseded_by in index.json and tables/lines.jsonl: the file of the newer line. So far that is each of the 12 original Center Finder lines, superseded by its polished copy. Use the newest line (one with no superseded_by) by default. Older versions are kept, never deleted, so a later polish adds a line and marks the one before it.

Since v2.6 (25 Oct 2026):

• Polished copies of my Center Finder lines on PHerc1299, PHercMAN5 and PHercMANB (2.4 um scans), beside the originals, which now carry superseded_by. The bucket publishes these scans' normal grids at level 2 only, so the polish ran in level-2 voxels (every 400 level-0 slices) and the points were scaled back to level 0 (polish_grid_level in index.json). The polish was only ever tested on 1.2 m scans at level 0, so on these three it is untested. Median moves: 1.3 mm (PHerc1299), 4.5 mm (PHercMAN5) and 1.6 mm (PHercMANB). The large move on PHercMAN5 means the original line sat off the innermost wrap there by villa's predicted sheets; check it before relying on either.

• New Center Finder lines for two rolled scrolls that had none: PHerc1667 (its 2.4 um scan; both ends pass the CT end check, which is untested at 2.4 um, so no end is marked usable) and PHerc0172 (its 7.91 um scan; both ends are flagged by the CT end check, so treat its ends with care). 35 scans in all.

• other_scans (index.json key, tables/other_scans.jsonl, files under other-scans/): 19 lines on other scans (18 in v2.6) of scrolls already here, each in its own scan's level-0 voxels. 16 are the scroll's default line carried to the other scan with the open-data catalog's own transform (metadata.json volume_transforms; source ending in -mapped, mapped_from names the original, transform_jackknife_um_median says how firmly the catalog's landmarks pin the transform there). That includes PHerc0139's line on volume 20260930100252, the remasked copy of the volume its line names (same voxel frame). 2 are new Center Finder lines: PHerc1451's 1.2 m scan (both ends pass) and PHerc1203's 2.4 um rescan (both pass, untested at 2.4 um). They are kept out of files so tools that pick a line by scroll keep getting the line for the scan the scrolls entry names. Mapped lines are not checked on their scan's CT by anyone.

• v2.6.1 (25 Oct 2026): PHerc0846A's 2.4 um rescan line added to other_scans (a new Center Finder line; its CT end check took longer than the v2.6 runs allowed and was run again with no time limit, same rule). Model output: no new line has been checked on CT by a person.

Since v2.7 (29 Oct 2026):

• Center Finder v2 lines (source finder-v2) on 37 scans of 33 scrolls. On the 13 scrolls that have a published person-made line, v2 started from the Center Finder fold that never saw that scroll; elsewhere it started from the Center Finder's own line. v2 moves the line slice by slice to the nearest point within 0.2 wraps of where villa's predicted sheet (its m7 normal grids) winds most, searching within 8 mm. It then smooths the line along z, with a running median over 12.6 mm and a Savitzky-Golay filter over 6.6 mm.

• The smoothing removes the sideways jumps the slice-by-slice polish leaves. Across the 34 lines on published grids, steps over 3 mm between neighboring points went from 571 to 38, and all 38 are within 20 mm of a line end.

• On the 13 scrolls with a person-made line, the smoothed v2 lines sit a median 1.47 mm from that line, with 64% of heights within 2 mm. Before smoothing it was 1.62 mm. Those 13 scrolls are development data for the finder, so this is not a fresh test.

• Where v2 replaces a line of mine, the older line now carries superseded_by. Person and bucket lines are never superseded, so tools that put people's lines first still do.

• PHerc0172, PHerc1667 and PHercParis3 have no published normal grids. For these three I ran villa's m7 surface model on one 192-slice layer per polish height, one height every 5.5 to 7.4 mm (with two gaps on PHerc0172), in a square 36 mm across round the line. I then made the grids with villa's vc_gen_normalgrids (grids: slabs in index.json). These three lines have fewer points than the rest, so where the scroll's axis leans they follow it in steps of a few mm, up to 13 mm on PHercParis3. This slab setup is untested.

• v2.8 (29 Oct 2026): Center Finder v2 lines for PHercParis4 and PHerc1451's 2.4 um scan, the two v2.7 left out. Paris 4's published grids come from villa's recto-2um model, not m7, so v2 is untested there. PHerc1451's line started from its 1.2 m Center Finder line, carried to the 2.4 um scan by a CT-only rotation registration. 39 v2 lines on 39 scans of 35 scrolls in all.

Model output: no v2 line or end has been checked on CT by a person.

Ten of these scans also have the whole villa lasagna field (cos, grad_mag, normals), one dataset each, such as https://huggingface.co/datasets/SharpMeow/pherc0175a-lasagna; all ten are listed at https://huggingface.co/spaces/SharpMeow/lasagna-fields