A config-driven, modular pipeline for preprocessing multiple Sign Language (SL) datasets. Supports two landmark extractors (MediaPipe Holistic and MMPose RTMPose3D) and two output modes (pose landmarks and video clips).

📖 New? See the Installation Guide to get started.

git clone https://github.com/balaboom123/signdata-slt.git
cd signdata-slt
python -m venv venv
source venv/bin/activate
pip install -r requirements.txt

MediaPipe works on CPU out of the box. MMPose requires a CUDA-capable GPU and additional dependencies -- see the Installation Guide for full setup instructions.

# Download YouTube-ASL videos, extract MediaPipe landmarks, normalize, and package into WebDataset shards
python -m signdata run configs/jobs/youtube_asl/mediapipe.yaml

# Extract MMPose landmarks from pre-downloaded How2Sign data (CUDA required)
python -m signdata run configs/jobs/how2sign/mmpose.yaml

# Override any config value from the command line (e.g. more workers, stop after extraction)
python -m signdata run configs/jobs/youtube_asl/mediapipe.yaml \
  --override processing.max_workers=8 stop_at=extract

Both modes produce WebDataset tar shards for efficient training data loading. See Pipeline Stages for detailed output formats and data shapes.

For paper-aligned preprocessing methodology, see Research-Aligned Preprocessing.

• Installation Guide -- base setup and MMPose GPU dependencies
• Architecture -- system design, registry, pipeline flow
• Configuration -- job/experiment layout and CLI overrides
• Pipeline Stages -- recipe stages and optional stages
• Datasets -- YouTube-ASL vs How2Sign setup
• Research-Aligned Preprocessing -- paper-aligned preprocessing notes

The MIT license in this repository applies to the code and documentation in this project. Use of external datasets, research artifacts, and upstream repos referenced above must comply with their original licenses and usage terms.

MIT -- see LICENSE.