Dataset Viewer
The dataset could not be loaded because the splits use different data file formats, which is not supported. Read more about the splits configuration. Click for more details.
Couldn't infer the same data file format for all splits. Got {NamedSplit('train'): ('json', {}), NamedSplit('test'): ('webdataset', {})}
Error code:   FileFormatMismatchBetweenSplitsError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

Cleaned GigaHands VITRA Linked Dataset

This directory is a cleaned copy of:

/home/chonghej/scratch/chonghej/VLA-HAND/datasets/vitra_gigahands_real_all_cam0_keypoints_mano_linked

The cleanup removes frame-level samples whose annotation frame id points past the actual decodable length of the linked RGB mp4.

What Was Cleaned

For each episode, the script checked:

  • video_decode_frame[frame_id]
  • actual decodable frame count from decord.VideoReader

Any sample with:

video_decode_frame[frame_id] >= video_frame_count

was removed from episode_frame_index.npz.

The source .npy episodic annotation files are hardlinked into this directory to avoid duplicating the full dataset storage.

Result

Split Episodes Before Episodes After Samples Before Samples After Samples Removed
gigahands_real_train 28,405 28,368 7,307,829 7,043,667 264,162
gigahands_real_test 1,496 1,495 383,606 368,585 15,021

No video open failures were found during cleanup.

Files

  • cleanup_report.json: top-level cleanup summary
  • Annotation/gigahands_real_train/cleanup_report.json: train split details
  • Annotation/gigahands_real_test/cleanup_report.json: test split details
  • Annotation/*/episode_frame_index.npz: cleaned sample indices

Statistics

The normalization statistics are copied from the original converted dataset for checkpoint compatibility. If this cleaned dataset is used for a new training run from scratch, recomputing train statistics on the cleaned split is recommended.

Usage

Download and extract the annotation archives with:

huggingface-cli download \
  --repo-type dataset \
  LeoJiangOR/vitra-gigahands-allcam0-keypoints-mano-cleaned \
  --local-dir datasets/vitra_gigahands_real_all_cam0_keypoints_mano_linked_cleaned_archives

mkdir -p datasets/vitra_gigahands_real_all_cam0_keypoints_mano_linked_cleaned
tar --zstd -xf \
  datasets/vitra_gigahands_real_all_cam0_keypoints_mano_linked_cleaned_archives/gigahands_real_train_annotations.tar.zst \
  -C datasets/vitra_gigahands_real_all_cam0_keypoints_mano_linked_cleaned
tar --zstd -xf \
  datasets/vitra_gigahands_real_all_cam0_keypoints_mano_linked_cleaned_archives/gigahands_real_test_annotations.tar.zst \
  -C datasets/vitra_gigahands_real_all_cam0_keypoints_mano_linked_cleaned

Then link the original GigaHands RGB videos:

mkdir -p datasets/vitra_gigahands_real_all_cam0_keypoints_mano_linked_cleaned/Video
ln -s /path/to/multiview_rgb_vids \
  datasets/vitra_gigahands_real_all_cam0_keypoints_mano_linked_cleaned/Video/GigaHands_root

The RGB videos are not included in this dataset repo. This release contains the converted VITRA annotations, cleaned frame indices, cleanup reports, manifests, and statistics.

Downloads last month
68