Dataset card: align with the JOCCH paper and repo contents

README.md

@@ -18,6 +18,9 @@ tags:
   - fabric
   - pattern-recognition
   - image
+  - jigsaw-puzzle
+  - archaeology
+  - benchmark
 size_categories:
   - 10K<n<100K
 source_datasets:
@@ -27,66 +30,112 @@ configs:
     data_files: "clean_dataset/clean_textiles_dataset.json"
 ---
 
-# TeXMET: Curated Textile Dataset from the Metropolitan Museum of Art
+# TEXMET: Curated Textile Dataset from the Metropolitan Museum of Art
+
+> **Paper:** Muhammad Hamza Zafar, Davit Gigilashvili, Jon Yngve Hardeberg. *TEXMET: A Deep Learning Benchmark for Archaeological Textile Reconstruction.* ACM Journal on Computing and Cultural Heritage, 2026 (Just Accepted). [doi.org/10.1145/3846195](https://doi.org/10.1145/3846195)
+>
+> **Code and full download:** [github.com/hamzafer/TEXMET](https://github.com/hamzafer/TEXMET)
 
 ## Overview
-**TeXMET** is a high-quality, manually curated dataset of textile and tapestry objects from the Metropolitan Museum of Art's Open Access collection. This dataset has been carefully cleaned, validated, and optimized for computer vision and deep learning applications.
+**TEXMET** is a high-quality, manually curated dataset of textile and tapestry objects from the Metropolitan Museum of Art's Open Access collection. It was built as a benchmark for jigsaw puzzle solving and archaeological textile reconstruction: models trained on TEXMET puzzles are evaluated on fragments of Viking Age textiles from the Oseberg ship burial. It is also usable for general computer vision work on historical textiles.
 
-## 🎯 TEXMET FINAL - CURATED DATASET
+## Key Facts
 - **Total Images**: 18,644 high-resolution images
-- **Unique Objects**: 1,697 textile/tapestry objects  
+- **Unique Objects**: 18,643 textile/tapestry objects (one image per object, except object 21466 with two; see Splits)
 - **Average Resolution**: 3.3 MP (1557 x 1631 pixels)
-- **Average File Size**: 1.35 MB per image
-- **Format**: 100% JPEG, 92% RGB mode
+- **Average File Size**: 1.42 MB per image
+- **Format**: 100% JPEG, 92% RGB, 8% grayscale
 - **Quality**: Hand-selected, duplicates removed, visually validated
 
+## Download the Full Dataset
+The images are fetched directly from the Met's Open Access servers with the script in the [TEXMET GitHub repository](https://github.com/hamzafer/TEXMET):
+
+```bash
+git lfs install                      # the metadata JSON is stored with Git LFS
+git clone https://github.com/hamzafer/TEXMET.git
+cd TEXMET
+pip install -r requirements.txt
+python download.py                   # writes to TEXMET_DATASET/
+```
+
+- `TEXMET_DATASET/images/` receives the primary image of every object (about 26.5 GB).
+- The script also fetches 17,120 additional views into `TEXMET_DATASET/additional_images/`. These are not part of the benchmark, apart from the one image noted under Splits, so plan for extra disk space.
+- The script uses the same `clean_textiles_dataset.json` as this repository (identical SHA-256).
+
+## Splits
+The [`splits/`](https://huggingface.co/datasets/hzafar/TEXMET/tree/main/splits) folder holds the exact train / validation / test partition used for every TEXMET experiment in the paper.
+
+| Split | Images |
+|---|---|
+| train | 14,915 |
+| val | 1,864 |
+| test | 1,865 |
+| **total** | **18,644** |
+
+- Image-level random shuffle, `random.seed(42)`, 80/10/10. The three lists are disjoint.
+- `train_files.txt`, `val_files.txt`, `test_files.txt`: one image path per line. The leading number of each filename is the Met `objectID`.
+- `texmet_splits.csv`: one row per image with `objectID`, `split`, `filename` and `imageURL`.
+- See [`splits/README.md`](https://huggingface.co/datasets/hzafar/TEXMET/blob/main/splits/README.md) for details.
+
+**Known quirk (one image).** The benchmark image set contains two images of object `21466` (its primary image and its fourth additional view, `21466_..._additional_4.jpg`, Met file `DP307213.jpg`), both in `train`, and no image of object `13562`. That is why 18,644 images cover 18,643 objects. When you use `download.py`, move `21466_..._additional_4.jpg` from `additional_images/` into `images/` and ignore `13562_..._primary.jpg`; the split lists then match the downloaded files exactly.
+
 ## Dataset Statistics
-- **Curation Method**: Manual validation using FiftyOne + CLIP embeddings
-- **Original Collection**: 27,373 objects → **Curated to 18,644 premium samples**
+- **Curation Method**: Manual validation in FiftyOne using CLIP ViT-B/32 embeddings, UMAP projection and LanceDB similarity search
+- **Original Collection**: 27,360 primary images → **curated to 18,644 images (68.1%)**; 8,716 removed (4,908 poor quality, 3,808 noise, including 2,567 duplicates)
 - **Quality Control**: Similarity analysis, duplicate removal, visual inspection
 - **Coverage**: Global textile traditions spanning 4000+ years
-- **Image Quality**: High-resolution, suitable for ML/CV applications
 
 ## Directory Structure
+This Hugging Face repository holds the metadata, the splits and a small image sample. The full image set (26.5 GB) is not stored here.
 ```
 TEXMET/
-├── clean_dataset/                   # 🎯 TEXMET FINAL CURATED DATASET
-│   ├── clean_textiles_dataset.json  # Main dataset (1,697 objects)
-│   ├── texmet_metadata.json         # Comprehensive metadata
-│   ├── images/                      # Primary images (~1,697 images)
-│   └── additional_images/           # Additional views (~17K images)
-├── FINAL_CORRECTED_MET_TEXTILES_DATASET/  # Original raw data (27,373 objects)
-│   └── objects_with_images_only/    # Complete raw collection
+├── clean_dataset/
+│   ├── clean_textiles_dataset.json  # Met metadata, one record per object (18,644 records)
+│   ├── texmet_metadata.json         # Summary statistics
+│   ├── images/                      # Sample only: 300 primary images
+│   └── additional_images/           # Sample only: 10 additional views
+├── splits/                          # Train / val / test lists used in the paper
 └── README.md                        # This file
 ```
 
+## Department Distribution
+TEXMET covers 14 Met curatorial departments.
+
+| Department | Images | % |
+|---|---:|---:|
+| European Sculpture and Decorative Arts | 8,516 | 45.7 |
+| Asian Art | 5,229 | 28.0 |
+| Islamic Art | 3,029 | 16.2 |
+| Other departments (11) | 1,870 | 10.0 |
+| **Total** | **18,644** | **100.0** |
+
 ## Data Quality & Curation Process
-1. **Initial Collection**: 27,373 objects from MET API (99.274% success rate)
+1. **Initial Collection**: 33,588 objects queried from the MET API, 27,373 with images; 55,007 images downloaded (99.95% success rate)
 2. **Filtering**: Focus on high-quality textile/tapestry objects
 3. **Manual Curation**: Visual inspection and validation using FiftyOne
 4. **Duplicate Removal**: CLIP embedding analysis + manual verification
 5. **Quality Control**: Removed low-quality, irrelevant, or damaged images
-6. **Final Validation**: **1,697 premium objects** with **18,644 images**
+6. **Final Validation**: **18,644 images** of **18,643 objects**
 
 ## Image Specifications
 - **Total Images**: 18,644 images
 - **Average Dimensions**: 1557 x 1631 pixels
-- **Average Aspect Ratio**: 1.14 (slightly portrait)
-- **Average File Size**: 1.35 MB
+- **Average File Size**: 1.42 MB
 - **Average Resolution**: 3.3 megapixels
 - **Format**: 100% JPEG
-- **Color Mode**: 92% RGB, 8% other modes
-- **Quality**: High-resolution, ML/CV ready
+- **Color Mode**: 92% RGB, 8% grayscale
 
 ## Research Applications
-- ✅ Textile pattern recognition and classification
-- ✅ Cultural heritage digitization and analysis
-- ✅ Style transfer and generative AI
-- ✅ Historical textile dating and attribution
-- ✅ Cross-cultural design studies
-- ✅ Computer vision algorithm development
-- ✅ Digital art and design inspiration
-- ✅ Large-scale image analysis and clustering
+- Jigsaw puzzle solving and fragment reassembly (the benchmark task in the paper)
+- Archaeological textile reconstruction
+- Textile pattern recognition and classification
+- Cultural heritage digitization and analysis
+- Style transfer and generative AI
+- Historical textile dating and attribution
+- Cross-cultural design studies
+- Computer vision algorithm development
+- Digital art and design inspiration
+- Large-scale image analysis and clustering
 
 ## Usage Examples
 ```python
@@ -98,16 +147,19 @@ with open('clean_dataset/clean_textiles_dataset.json', 'r') as f:
     data = json.load(f)
 
 df = pd.DataFrame(data)
-print(f"TeXMET Final: {len(df)} curated objects")
+print(f"TEXMET Final: {len(df)} curated images")
 
-# Load metadata
+# Load summary statistics
 with open('clean_dataset/texmet_metadata.json', 'r') as f:
     metadata = json.load(f)
-print(f"Total images: {metadata['image_statistics']['total_images']}")
+stats = metadata['dataset_statistics']
+print(f"Objects with a primary image: {stats['objects_with_primary_images']}")
+print(stats['department_distribution'])
 ```
 
 ## Version History
-- **TeXMET Final v1.0** (July 2025): **1,697 curated objects, 18,644 images**
+- **September 2026**: added `splits/`, restored the complete `texmet_metadata.json`, card aligned with the JOCCH paper
+- **TEXMET Final v1.0** (July 2025): **18,644 curated images of 18,643 objects**
 - **Raw Collection** (July 2025): 27,373 objects collected from MET API
 
 ## Technical Specifications
@@ -115,24 +167,25 @@ print(f"Total images: {metadata['image_statistics']['total_images']}")
 - **Image Format**: 100% JPEG
 - **Resolution**: High-resolution (average 3.3 MP)
 - **File Organization**: Git LFS for large files
-- **Compatibility**: Ready for PyTorch, TensorFlow, OpenCV
 
 ## Citation and Credits
 
-### TeXMET Dataset Citation
+### TEXMET Citation
+If you use TEXMET, please cite the paper:
 ```bibtex
-@dataset{texmet2025,
-  title={TeXMET: Curated Textile Images from the Metropolitan Museum of Art},
-  author={HAMZA},
-  year={2025},
-  publisher={Thesis Project},
-  version={1.0},
-  note={Manually curated dataset: 1,697 objects, 18,644 images}
+@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},
+  publisher = {Association for Computing Machinery},
+  doi       = {10.1145/3846195},
+  note      = {Just Accepted}
 }
 ```
 
 ### Original Data Source
-This dataset (TeXMET) is a curated subset of textiles and tapestries from The Metropolitan Museum of Art's Open Access collection, compiled for research purposes. To the best of our knowledge, this is the first comprehensive compilation of MET's textile and tapestry collections in this format, manually curated and optimized for computer vision applications.
+This dataset (TEXMET) is a curated subset of textiles and tapestries from The Metropolitan Museum of Art's Open Access collection, compiled for research purposes. To the best of our knowledge, this is the first comprehensive compilation of MET's textile and tapestry collections in this format, manually curated and optimized for computer vision applications.
 
 ```bibtex
 @misc{MET_OpenAccess_2024,
@@ -154,20 +207,12 @@ This dataset (TeXMET) is a curated subset of textiles and tapestries from The Me
 - **Data Source**: The Metropolitan Museum of Art Open Access Initiative
 - **Curation Tools**: FiftyOne, CLIP embeddings, manual validation
 - **API Access**: MET Museum Collection API
-- **Project**: TeXMET - Textile Analysis Thesis Project
+- **Project**: TEXMET benchmark (ACM JOCCH 2026)
 - **Created**: July 2025
 
-## Performance & Quality Metrics
-- **Image Quality**: 100% JPEG, high resolution
-- **Duplicate Removal**: CLIP similarity + manual verification
-- **Manual Validation**: Visual inspection of all selected objects
-- **Research Ready**: Optimized for ML/CV applications
-
 **Note**: This is a research compilation and curation of existing open access data. All rights and credits belong to The Metropolitan Museum of Art. We thank the MET for making their collection freely available for research and education.
 
 ---
-**🎯 TeXMET Final: 1,697 Premium Objects, 18,644 High-Quality Images**
-Total Dataset Size: 26.5 GB.
-**Ready for Computer Vision and Machine Learning Applications!**
+**TEXMET Final:** 18,644 images of 18,643 objects, 26.5 GB.
 
-*Last Updated: July 8, 2025*
+*Last Updated: September 15, 2026*

clean_dataset/texmet_metadata.json

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:9e7d8b2ee71550f6e01eab552122c9268e6eddd943f4de96cdbf02a118e3fce8
-size 765
+oid sha256:c12224c1758d7c4ba3e75fe0661da524271d10bc8dbb69ed5f9485d966ff4b71
+size 3599

splits/README.md

@@ -23,6 +23,7 @@ Image-level random shuffle with `random.seed(42)`, 80/10/10. The three lists are
 
 ## Notes
 
-- Each image is the primary image of a distinct museum object, so the split is also object-level, with one exception: object `21466` contributes two images (its primary image and an additional view, `..._additional_4.jpg`), and both are in `train`, so no object crosses splits. Its `imageURL` is left empty in the CSV because the additional-view index is ambiguous. This gives 18,643 unique objects for 18,644 images.
+- Each image is the primary image of a distinct museum object, so the split is also object-level, with one exception: object `21466` contributes two images (its primary image and an additional view, `..._additional_4.jpg`), and both are in `train`, so no object crosses splits. Its `imageURL` is the Met file `DP307213.jpg` (checked byte-for-byte against the benchmark copy). Object `13562` is in `clean_textiles_dataset.json` but has no image in the benchmark. This gives 18,643 unique objects for 18,644 images.
+- With `download.py`, move `21466_..._additional_4.jpg` from `additional_images/` into `images/` and ignore `13562_..._primary.jpg`; the three lists then match the downloaded filenames exactly.
 - Distinct Met objects can come from related historical textiles (for example several catalogued fragments of one source fabric); no metadata-level grouping was applied for such cases.
 - Images are CC0 (Metropolitan Museum of Art Open Access). Use the `download.py` script in https://github.com/hamzafer/TEXMET to fetch them.

splits/texmet_splits.csv

@@ -12994,7 +12994,7 @@ objectID,split,filename,imageURL
 66855,train,66855_Piece_primary.jpg,https://images.metmuseum.org/CRDImages/as/original/DP11347.jpg
 224187,train,224187_Sample_primary.jpg,https://images.metmuseum.org/CRDImages/es/original/117292.jpg
 463277,train,463277_Textile_Fragment_primary.jpg,https://images.metmuseum.org/CRDImages/md/original/133182.jpg
-21466,train,21466_Hanging_depicting_a_European_conflict_in_South_Ind_additional_4.jpg,
+21466,train,21466_Hanging_depicting_a_European_conflict_in_South_Ind_additional_4.jpg,https://images.metmuseum.org/CRDImages/es/original/DP307213.jpg
 221874,train,221874_Leda_and_the_Swan_primary.jpg,https://images.metmuseum.org/CRDImages/es/original/DP-34105-001.jpg
 227588,train,227588_Music_primary.jpg,https://images.metmuseum.org/CRDImages/es/original/DT815.jpg
 216875,train,216875_Piece_primary.jpg,https://images.metmuseum.org/CRDImages/es/original/6783.jpg