9 um ink detection across scrolls: scan-type transfer and self-training
Checkpoints and small data files for reading unseen Herculaneum scrolls from their native 8.6-9.4 um CT scans with
the flat 9 um ink recipe of ScrollPrize/villa
(vesuvius/src/vesuvius/ink_detection/configs/aligned21_hybrid_3d2d.json). Method and code: docs/ink_cross_scroll.md,
training/scan_transfer.py and pseudo_labels.py in the accompanying villa pull request.
Results on an unseen scroll (PHerc 0139 held out)
All loso0139/ models were trained on PHerc 1667, Paris 4 and 0814 only (25k steps, batch 64). Mean per-pixel ROC-AUC on
PHerc 0139, forward direction, inside the open-data supervision masks:
| checkpoint | native 9.362 um (5 segments), plain | native, mirror TTA | pooled renders (9 segments) |
|---|---|---|---|
base_seed42.pth |
0.743 | 0.779 | 0.821 |
base_seed43.pth / base_seed44.pth |
0.783 / 0.763 | 0.800 / 0.792 | 0.822 / 0.831 |
base_seed42_78k_steps.pth (recipe length) |
0.754 | 0.780 | 0.780 |
scan_transfer_seed42.pth (+ scan-type transfer) |
0.785 | 0.797 | 0.826 |
selftrain_safer_seed42.pth (baseline recipe + self-training on all 5 native 0139 segments; transductive, see below) |
0.804 | 0.813 | 0.835 |
average of the four base_* (the self-training teacher) |
0.799 | 0.808 | 0.850 |
selftrain_split_seed42.pth (baseline recipe + self-training on w035 w040 w044 only; 3 of the 5 scored segments adapted) |
0.812 | 0.820 | 0.838 |
Scan transfer over three seeds (42 / 43 / 44, each paired with the baseline of the same seed; only seed 42's checkpoint is published): native plain 0.785 / 0.789 / 0.786, +0.024 over the baselines (SE 0.003 across segments); with mirror TTA 0.797 / 0.800 / 0.806, +0.011 (SE 0.005); pooled 0.826 / 0.825 / 0.832, no loss. Seed 42 is the most favourable pairing.
Split test (the First Letters setting): on w039 and w041, which never contributed a pseudo-label,
selftrain_split_seed42.pth scores 0.770 and 0.841 with mirror TTA (teacher 0.778 / 0.828, base_seed42.pth 0.735 / 0.803).
production/: the same recipe with the scan transfer, trained on all four labelled scrolls (1667, Paris 4, 0814, 0139),
seeds 42, 43 and 44; for inference on unread scrolls (average the three, with mirror TTA).
Replaced on 2026-11-13 by retrains with skip_datasets: ["native9"], so the already-native 0139 group is no longer
converted; the earlier files, which converted it, remain in this repository's history. On two scrolls none of these
models saw (seeds 42 / 43 / 44, ensemble in brackets): PHerc 0841 human labels 0.753 / 0.774 / 0.781 (0.782), earlier
files 0.758 / 0.769 / 0.777 (0.780); PHerc 0343P camera truth 0.748 / 0.741 / 0.759 (0.759), earlier files 0.732 /
0.719 / 0.742 (0.740). Seed 44 was trained on a different GPU with the same code and config.
selftrain_safer_seed42.pth was trained with pseudo-labels on the same five native 0139 segments it is scored on
(no true labels used): this measures adaptation to a target scroll, as one would on an unread scroll. On the four pooled
0139 segments that were never pseudo-labelled it scores 0.830. A model trained with 0139's labels scores 0.99+.
Files
loso0139/*.pth: inference checkpoints (model weights + training config; optimizer state removed). Load withpython -m vesuvius.ink_detection.inference.infer SURFACE_VOLUME.zarr CKPT.pth OUT.tif --direction both --tta-mirror. The scan-transfer checkpoints carry theirscan_transferblock in the saved config (added toscan_transfer_seed42.pthon 2026-11-13, weights unchanged); the self-training checkpoints use the baseline recipe.configs/*.json: the training configs (/path/to/ink_9um= the ink_9um label dataset).loso0139_selftrain_safer.jsonadds a dataset group0139pof pseudo-labelled native 0139 segments built withvesuvius.ink_detection.pseudo_labels(--background max --margin 8, teachers = the fourbase_*models with mirror TTA).scan_transfer/pooled_to_native_pherc0814.json: the pooled -> native kernel (fitted on PHerc 0814 segment 46527, R^2 0.707 vs 0.580 for resampling alone on held-out spatial blocks), shipped in villa asbuiltin:pooled_to_native_pherc0814.transforms/PHerc0846A_20260319102732-to-20250728152254/transform.json: an affine from the PHerc 0846A 2.403 um / 77 keV volume to its 9.362 um / 113 keV volume, in the organisers' transform format (own fit, held-out NCC 0.916 at ~75 um; renders through it match native renders at plane NCC 0.83-0.85). Not an official transform.
Data and licence
Trained on Vesuvius Challenge open data (scrollprize.org/data), released under CC BY-NC 4.0; these weights are released under the same licence. The flat 9 um recipe, its labels and the inference code are the Vesuvius Challenge team's (ScrollPrize/villa); this release adds the scan transfer, the self-training procedure and their evaluation. Data citation: G. Angelotti, S. Parsons, S. Johnson, E. R. Dal Pra, J. Rudolph, P. Tafforeau, A. Mirone, P. Henderson, H. Schilling, F. McDonald, D. Josey, Y. Nader, C. S. Parker, W. B. Seales. Vesuvius Challenge - CT Scans of Herculaneum Papyri. Vesuvius Challenge. Includes data derived from the EduceLab-Scrolls dataset: Parsons, S., Parker, C. S., Chapman, C., Hayashida, M., Seales, W. B. (2023). EduceLab-Scrolls: Verifiable Recovery of Text from Herculaneum Papyri using X-ray CT. arXiv:2304.02084.