GAP-fast

GAP-fast is a storage-repacked version of the official GAP benchmark, created by Ofir Itzhak Shahar, Gur Elkin, and Ohad Ben-Shahar. This archive was prepared by Changxin Ye to make GAP training and evaluation with RG-LNS easier to reproduce.

What was changed

The original GAP HDF5 files use gzip compression and chunked storage. Random reads can therefore decompress data belonging to other puzzles. GAP-fast rewrites the puzzle and index-label datasets using contiguous, uncompressed HDF5 storage to reduce this overhead.

The conversion preserves the puzzle pixel values, labels, sample order, and train/validation/test splits. It does not regenerate puzzle fragments. Original coordinate-label files are omitted because coordinates can be recovered from index labels as row = label // grid_size and col = label % grid_size.

Each split includes repack_info.json with conversion details, sampled puzzle-equality checks, full index-label equality checks, and read-speed measurements. These checks do not constitute an exhaustive comparison of every puzzle pixel in the dataset.

Download and extraction

Download GAP_fast.zip from the Files and versions tab. The archive is approximately 13.5 GB and requires about 44.6 GB of disk space when extracted. Allow additional space if you retain the archive.

Extract the archive into the datasets/ directory of RG-LNS. The ZIP already contains the top-level GAP_fast/ folder, so the final layout should be:

RG-LNS/
└── datasets/
    └── GAP_fast/
        ├── GAP-3/
        │   ├── train/
        │   ├── val/
        │   └── test/
        ├── GAP-5/
        │   ├── train/
        │   ├── val/
        │   └── test/
        ├── README.md
        └── repack_gap_hdf5.py

Each split contains:

puzzles.h5
labels_indices.h5
metadata.json
metadata_original.json
repack_info.json

The RG-LNS GAP loader uses datasets/GAP_fast/ by default. See the RG-LNS README for environment, checkpoint, training, and evaluation instructions. The included repack_gap_hdf5.py can also regenerate the storage layout from the original GAP files.

Provenance and license

The images, fragment masks, puzzle splits, and labels originate from Ofirish/GAP. The original authors retain credit for the GAP benchmark; this repository documents the storage conversion and provides a convenient archive.

The data follows the original dataset's CC BY-NC 4.0 license. Preserve attribution, identify the storage conversion, and observe the non-commercial terms when using or redistributing the data. Refer to the original dataset card for its intended academic research use, benchmark protocols, and limitations.

Citation

Please cite the original GAP paper when using the benchmark:

@InProceedings{Shahar_2026_CVPR,
    author    = {Shahar, Ofir Itzhak and Elkin, Gur and Ben-Shahar, Ohad},
    title     = {The Missing GAP: From Solving Square Jigsaw Puzzles to Handling Real World Archaeological Fragments},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2026},
    pages     = {3186--3196}
}