Eurographics GCH 2026

Barcelona, Spain

TLDA (Too Long Didn't Attend)

I presented TEXMET: A Deep Learning Benchmark for Archaeological Textile Reconstruction (with D. Gigilashvili and J. Y. Hardeberg) at the Eurographics Workshop on Graphics and Cultural Heritage, UPC Campus Nord. The paper is published in the ACM Journal on Computing and Cultural Heritage (JOCCH), Visual Heritage special issue. It started as my master's thesis.


The problem

Old textiles often survive only as small, faded, broken pieces. The Viking Age Oseberg ship burial in Norway (834 CE) left hundreds of such fragments, and fitting them back together by hand is slow, careful work.

AI models that solve image jigsaw puzzles already exist, but they were trained on everyday photos, not on woven fabric.

What we built

  • TEXMET: 18,644 images of historical textiles from the Metropolitan Museum of Art's open collection (CC0).
  • Curation: 55,007 images → 27,360 primary images → CLIP ViT-B/32 embeddings + UMAP + LanceDB to find duplicates and outliers → human review → 18,644.
  • Fixed splits: 14,915 / 1,864 / 1,865 (train / val / test), so everyone tests on the same images.
  • The task: cut each image into a 3×3 grid of square pieces, shuffle them, and measure how often a model puts every piece back in the right place.

Results

  • We tested three modern puzzle-solving models. Trained only on everyday photos, they do poorly on fabric.
  • One extra round of training on TEXMET changes that: the best model, DiffAssemble, solves 85.89% of textile puzzles.
  • On 45 heavily damaged Oseberg fragments, its success rate rose from 26.67% to 64.44%.
  • The lesson: AI tools for cultural heritage need training data that looks like the real material. The same approach could help with pottery, wall paintings and manuscripts.

Questions from the audience

Is the process fair, transparent and reproducible?

For using the benchmark, yes.

  • The source is The Met Open Access (CC0), so anyone can download the same originals.
  • The curated images and the full Met metadata per object are on Hugging Face; the download script and requirements are on GitHub.
  • The splits are fixed, and every curation step is in the paper with its count.

Caveat: the final tagging (poor quality or noise) was interactive human review in FiftyOne, not a fixed rule. Re-running the pipeline from scratch gets close but not identical, so reproducing the benchmark means using the released curated set. Since the talk, the list of removed objects (with their Met metadata) is published too (curation/removed_objects.json). The curation tag behind each removal (poor quality or noise) isn't published yet; that's the step that would make the human decisions fully transparent.

How would you build a similar dataset from another source, and how many images do you need?

  1. Pick an open-access collection with an API (e.g. the Rijksmuseum, the Smithsonian, the V&A, the Cleveland Museum of Art).
  2. Harvest by category (textiles, tapestries) and keep one main image per object.
  3. Curate: CLIP embeddings → UMAP map → similarity search for duplicates and outliers → human review.
  4. Fix a train / val / test split, then generate the puzzles.

How many: we used about 15,000 training images and that was enough to close most of the gap. We haven't tested the minimum; that's a good open question. A follow-up study: fine-tune on 1k / 5k / 10k / 15k images and plot accuracy against size.