Dataset Card for TEXMET

Dataset Summary

TEXMET is a curated dataset of 18,644 historical textile images from The Metropolitan Museum of Art Open Access collection: tapestries, woven and embroidered fabrics, lace and fragments spanning many periods and cultures. It was built as, to our knowledge, the first textile-specific benchmark for deep learning puzzle solving, as a step toward reconstructing fragmented archaeological textiles such as the Viking Age Oseberg tapestries.

In the accompanying paper, fine-tuning the DiffAssemble puzzle solver on TEXMET raised its accuracy on real Oseberg fragments from 26.7% to 64.4%.

from datasets import load_dataset

ds = load_dataset("hzafar/TEXMET")

Supported Tasks

• Puzzle solving / reassembly: images are cut into tiles (3×3 in the paper), shuffled, and models predict the original arrangement.
• Domain adaptation: pre-training or fine-tuning vision models on historical textiles before applying them to archaeological material.
• The Met metadata (department, culture, period, medium) also supports textile classification and retrieval.

Dataset Structure

Data Instances

{
  "image": <PIL.JpegImageFile>,
  "objectID": 69980,
  "filename": "69980_Fragment_of_Galloon_primary.jpg",
  "title": "Fragment of Galloon",
  "objectName": "Galloon fragment",
  "department": "Asian Art",
  "culture": "Korea",
  "period": "Joseon dynasty (1392–1910)",
  "objectDate": "late 17th century",
  "objectBeginDate": 1671,
  "objectEndDate": 1699,
  "medium": "Satin",
  "dimensions": "46 1/4 x 5 in. (117.47 x 12.70 cm)",
  "classification": "Textiles-Woven",
  "creditLine": "Gift of Millia Davenport, 1953",
  "objectURL": "https://www.metmuseum.org/art/collection/search/69980",
  "primaryImage": "https://images.metmuseum.org/CRDImages/as/original/DP14339.jpg"
}

Data Fields

• image: the image, stored as the original JPEG bytes downloaded from The Met (unchanged).
• objectID: Met object ID.
• filename: original file name.
• title, objectName, department, culture, period, objectDate, objectBeginDate, objectEndDate, medium, dimensions, classification, creditLine: Met catalog fields.
• objectURL, primaryImage: links to the Met record and the source image.

Data Splits

Split	Images
train	14,915
validation	1,864
test	1,865
total	18,644

These are the exact splits used in the paper (image-level shuffle, random.seed(42), 80/10/10). Each image is the primary image of a distinct Met object, so the split is also object-level (one exception, see Other Known Limitations).

Dataset Creation

Curation Rationale

Puzzle-solving models are trained on natural photographs (ImageNet) and largely struggle on historical textiles, and no public textile dataset existed for this task. TEXMET fills that gap with a large, openly licensed set of curated textile images.

Source Data

33,588 textile and tapestry objects were queried through the Met Open Access API; 27,373 had images, and 55,007 images were downloaded (99.95% success). One primary image per object was kept: 27,360 images (13 objects had no downloadable primary image).

Curation Process

1. Embedding: CLIP ViT-B/32 embeddings (512-d) for every image.
2. Mapping: UMAP projection to reveal clusters and outliers.
3. Search: LanceDB nearest-neighbor search to find look-alikes and duplicates.
4. Review: interactive human review in FiftyOne.

8,716 images were removed: 4,908 as poor quality (for example armor or weapons filed as textiles) and 3,808 as noise (single threads, near-empty images), including 2,567 duplicates. 18,644 images remain (68.1% of the primary images). The final tagging was a human decision, so re-running the pipeline gets close but not identical; use this release to reproduce the benchmark. The metadata of all removed objects is published in curation/removed_objects.json.

Additional Files

Path	Contents
clean_dataset/clean_textiles_dataset.json	Full Met metadata for the 18,644 objects, incl. URLs of 17,120 additional views (not part of the benchmark)
splits/	Split lists and texmet_splits.csv with the Met image URL of every file
curation/removed_objects.json	Met metadata of the 8,729 removed objects (13 without a downloadable image, hence 8,716 removed images)
curation/processing_log.csv	Per image: original size, mask area and the inscribed and regular crop sizes used in the paper's experiments

Considerations for Using the Data

Other Known Limitations

• Puzzles in the paper are regular 3×3 grids cut from intact images; real fragments are irregular and degraded.
• Object 21466 contributes two images (its primary image and one additional view), both in train, so the 18,644 images come from 18,643 objects. Object 13562 appears in clean_textiles_dataset.json but not in the splits.
• Distinct Met objects can be cataloged fragments of the same source fabric; no grouping was applied for such cases.

Additional Information

Dataset Curators

Muhammad Hamza Zafar, Jon Yngve Hardeberg and Davit Gigilashvili, Colourlab, Norwegian University of Science and Technology (NTNU).

Licensing Information

All images and metadata are CC0 (public domain) under The Met Open Access policy. Credit to The Metropolitan Museum of Art is appreciated.

Citation Information

@article{zafar2026texmet,
  author  = {Zafar, Muhammad Hamza and Gigilashvili, Davit and Hardeberg, Jon Yngve},
  title   = {{TEXMET}: A Deep Learning Benchmark for Archaeological Textile Reconstruction},
  journal = {ACM Journal on Computing and Cultural Heritage},
  year    = {2026},
  doi     = {10.1145/3846195},
  eprint  = {2611.04734},
  archivePrefix = {arXiv}
}