Modelpublic

scrollprize/ink_3d_dino_guided

sourceHugging Facemitupdated 2mo agoView on Hugging Face

0likes17kdownloads

Model Card

PHerc. Paris 4 — volumetric 3D ink segmentation (DINO-guided, self-distilled)

Segments ink directly in 3D in micro-CT of PHerc. Paris 4, where ink is volumetrically visible. This model provides the independent volumetric validation of surface-conditioned ink recovery reported in "Complete virtual unwrapping and reading of a rolled Herculaneum papyrus" (Angelotti et al., arXiv:2606.29085, 2026).

Model details


Architecture	`vesuvius` `NetworkFromConfig` nnU-Net-style 3D residual-encoder U-Net (~142 M params), single sigmoid ink head
Input	1-channel CT, 256³ patches, percentile min-max normalisation, full-3D mode (projection half-thickness 3 voxels)
Optimisation	SGD + Nesterov, cosine LR (base 1e-2, warmup 5000), bf16, Dice+BCE with 0.1 label smoothing, weight EMA
This checkpoint	step 78,000 · W&B run `4b07qv8p` (`ps256_3d_bcedice_dinoguided_paris4_v3_fullsup`)
Weights	`model` (raw) and `ema_model` (EMA — recommended for inference)

DINO-guided supervision + self-distillation

Trained in stages: (1) teacher on 8-scroll surface-conditioned ink labels (PHerc. Paris 4, 0139, 0500P2, 0814, 0841, 1667, MAN5, 9B); (2) a DINO-guided student using dense ink-likeness from the 3D DINOv2 representation model `scrollprize/dinovol_v2_ps8_with_paris4_352500`; (3) + background masking; (4) self-distillation. DINO guidance compares each 864-D patch token to a reference ink embedding (avg_ref_embedding.npy, the L2-normalised mean of 256 expert-clicked tokens stored in recorded_embeddings*.npy) at threshold τ = 0.5. Full configuration is in Supplementary Table 4 of the paper.

Files

—ckpt_78k_fullsup.pth — checkpoint with model + ema_model (use ema_model for inference). Training/inference config embedded under config and mirrored in config.json.
—config.json — training/inference configuration.
—avg_ref_embedding.npy — (864,) reference ink embedding for DINO guidance.
—recorded_embeddings.npy, recorded_embeddings_2.npy — (128, 864) each; the expert-clicked DINO tokens (256 total) the reference embedding is averaged from.

How to load

python

import torch, numpy as np
ck    = torch.load("ckpt_78k_fullsup.pth", map_location="cpu", weights_only=False)
state = ck["ema_model"]                    # recommended
cfg   = ck["config"]                       # == config.json
ref   = np.load("avg_ref_embedding.npy")   # (864,) reference ink embedding
# Build the network with the vesuvius package (NetworkFromConfig(cfg)), then load_state_dict(state)

The vesuvius package and inference code are in <https://github.com/ScrollPrize/villa>.

Related models

—3D representation (guidance): `scrollprize/dinovol_v2_ps8_with_paris4_352500`
—2D ink detector + label ablations: `pherc1667-ink-detection-ablation` collection

Citation

bibtex

@misc{angelotti2026unwrapping,
  title         = {Complete virtual unwrapping and reading of a rolled Herculaneum papyrus},
  author        = {Angelotti, Giorgio and others},
  year          = {2026},
  eprint        = {2606.29085},
  archivePrefix = {arXiv},
  primaryClass  = {eess.IV},
  doi           = {10.48550/arXiv.2606.29085}
}

License

MIT — released by the Vesuvius Challenge. The underlying tomographic data are distributed under CC BY-NC 4.0 (see the data links above).