WhynotGPT/ReInterHand-JPEG-Q95

Re:InterHand JPEG Q95

An unofficial JPEG storage derivative of Re:InterHand by Gyeongsik Moon and colleagues, published at NeurIPS 2023 Datasets and Benchmarks. The original dataset and this derivative retain CC BY-NC 4.0 terms. See the license. This repository is maintained by WhynotGPT, not the original dataset authors.

Upload status: check upload_status.json. Do not treat this repository as a complete download until complete is true and all expected shards are present.

Contents

• All 10 captures, 1,526,795 RGB images from the local pretraining copy.
• RGB JPEG payload: 244,755,056,430 bytes (244.8 GB, approximately 228 GiB), plus annotation files and tar headers.
• Four variants: Mugsy third-person and ego cameras, each with per-frame and per-segment environment maps.
• Unmodified source frame_list.txt, frame_list_orig.txt, Mugsy camera parameters, ego per-frame camera parameters and truncation ratios, and orig_fits/{left,right}/Keypoints/*.json.

RGB was converted from the original PNGs using Pillow image.convert("RGB").save(..., format="JPEG", quality=95, optimize=True) with default chroma subsampling. This is lossy compression, without resizing or cropping. Frame IDs, camera coordinates and GT keypoints were not changed. Image suffixes are .jpg instead of .png; loaders must use the new suffix.

This is the RGB/keypoint pretraining storage version. It does not include segmentation masks, source meshes, MANO fits, or MANO model binaries. For those, use the official downloads linked on the original dataset page. In particular, orig_fits/Keypoints are the source rendered-geometry keypoints, not our own new MANO fits and not the separate keypoints_orig capture annotations.

Splits and geometry

All official train and test captures are included. The local pretraining pool uses all captures; results from such training must not be reported as held-out performance under the original Re:InterHand test protocol. Use the official capture split if an independent test is required.

The source 3D keypoints and camera translations use millimeters. The retained keypoint order follows the official orig_fits/Keypoints definition. Consult the official directory and geometry documentation for keypoint mapping and camera projection conventions.

catalog.json provides the frame/camera sampling inventory with a portable root of ReInterHand. dataset_manifest.json reports capture/variant counts.

Download and extraction

No account or access approval is required to read this public repository. Install huggingface_hub, then download:

hf download WhynotGPT/ReInterHand-JPEG-Q95 --repo-type dataset --local-dir reih_q95_download
python reih_q95_download/verify_and_extract.py --download-dir reih_q95_download --extract-to /path/to/pretrain

The verification utility uses Python 3.9+ and the standard library. It verifies each tar's size and SHA-256 before extraction and refuses to overwrite existing files. Use an empty destination for the restored ReInterHand/ directory. Allow room for both the downloaded shards and extracted data.

Shards are split by capture and variant, targeting at most 4 GB of source data per shard. Metadata has separate shards. Each member starts with ReInterHand/<capture>/, so extracting all shards preserves the original directory structure. shards_manifest.jsonl and SHA256SUMS cover the new tar files; original PNG checksums do not apply to JPEGs.

Citation

@inproceedings{moon2023reinterhand,
  title={A Dataset of Relighted {3D} Interacting Hands},
  author={Moon, Gyeongsik and Saito, Shunsuke and Xu, Weipeng and Joshi, Rohan
    and Buffalini, Julia and Bellan, Harley and Rosen, Nicholas and Richardson,
    Jesse and Mize Mallorie and Bree, Philippe and Simon, Tomas and Peng, Bo
    and Garg, Shubham and McPhail, Kevyn and Shiratori, Takaaki},
  booktitle={NeurIPS Track on Datasets and Benchmarks},
  year={2023}
}