wlasl_reduced / README.md
jherng's picture
Upload folder using huggingface_hub
8067c4e verified

Reduced WLASL Dataset

This dataset is a task-specific reduced version of the WLASL dataset, constructed for American Sign Language (ASL) recognition experiments.

Contents

  • videos/
    Video clips organized by gloss label.

  • metadata.csv
    Per-sample metadata including:

    • file path
    • gloss label
    • fps (after normalization, if applied)
    • video resolution
    • normalized bounding box coordinates
  • gloss_map.json
    Mapping from gloss labels to integer class IDs.

Dataset Construction

The dataset was generated from the original WLASL dataset using a custom CLI preprocessing script with the following criteria:

  • Glosses restricted to a selected subset
  • Samples without bounding boxes were excluded
  • Videos optionally normalized to a fixed FPS
  • Original videos preserved (no cropping applied)
  • Bounding boxes stored as metadata for downstream processing

Command Used

The dataset was generated using the following command:

python .\src\sign_language_model\scripts\build_reduced_wlasl.py --wlasl-root .\data\WLASL\ --output-root .\data\wlasl_reduced --glosses before,cool,thin,go,drink,help,computer,cousin,who,bowling,trade,bed,accident,tall,thanksgiving,candy,short,pizza,man,no,wait,good,bad,son,like,doctor,now,find,you,thank you,please,hospital,bathroom,me,i --target-fps 24 --dry-run

Notes

  • Bounding boxes are stored in normalized coordinates (0-1) with top-left origin.
  • FPS normalization (if applied) uses frame dropping (no interpolation).