features_kps/README.md

Holistic Keypoint Features

This directory contains pre-extracted MediaPipe keypoint features derived from the reduced WLASL dataset.

Feature Description

• Modality: MediaPipe Holistic
• Keypoints included:
  • Pose (33 points): (x, y, z) (4 dims)
  • Left hand (21 points): (x, y, z) (3 dims)
  • Right hand (21 points): (x, y, z) (3 dims)
• Frames per video: 32 (fixed length)

Each feature file is stored as a NumPy array with shape: (C, T, V) = (3, 32, 75), where:

• C: Coordinate dimensions (x, y, z)
• T: Number of frames (fixed to 32)
• V: Number of keypoints (75 total)

Preprocessing Details

• Videos are cropped using the provided bounding boxes.
• Bounding boxes are expanded by 10% on all sides before cropping.
• Keypoints are extracted from all frames in the video.
• Frames without detected landmarks are discarded.
• From the remaining frames:
  • If at least 32 frames are available, 32 frames are uniformly sampled.
  • If fewer than 32 frames are available, last-frame padding is applied.

Directory Structure

data/wlasl_reduced/features_kps/
├── <gloss_label>/
│   ├── <video_id>.npy  # shape: (3, 32, 75)
│   ├── ...
├── warnings.txt # Optional: Warnings during extraction
└── README.md

Command Used

The features were generated using the following command:

python .\src\sign_language_model\scripts\extract_holistic_keypoints.py --dataset-root .\data\wlasl_reduced\ --output-root .\data\wlasl_reduced\features_kps --crop-to-bbox --num-workers 4