Siruyy/asl-static-landmarks-v1

ASL Static Letter Recognizer (Real-Time)

Real-time ASL fingerspelling recognizer built with MediaPipe + TensorFlow.

This project recognizes 24 static ASL letters from webcam input.

• ✅ Included: static letters (e.g., A, B, C, ...)
• ❌ Excluded: J, Z (motion-based letters)

Current model status

• Final static model: models/asl_static_model.h5
• Label encoder: models/static_label_encoder.pkl
• Input features: 86 engineered features from 21 hand landmarks
• Reported accuracy: 92.74% overall, 94.28% per-letter average

Project structure

aisl/
├── src/
│   ├── hand_detector.py
│   ├── feature_engineer.py
│   ├── data_collector.py
│   ├── data_processor.py
│   ├── train_static_only.py
│   ├── analyze_model.py
│   └── predictor_static.py
├── models/
│   ├── asl_static_model.h5
│   └── static_label_encoder.pkl
├── data/
│   ├── processed/
│   ├── processed_v2/
│   └── raw/ (ignored by git)
├── requirements.txt
└── README.md

Setup

1. Create and activate a virtual environment.

2. Install dependencies:

   pip install -r requirements.txt

Run live prediction

From project root:

python3 src/predictor_static.py

Controls

• q: quit
• + / -: increase/decrease confidence threshold

Training workflow (optional)

1. Collect data (raw landmarks)
2. Engineer features and prepare dataset
3. Train static model
4. Analyze confusion and per-letter performance

Primary scripts:

• src/collect_more_static.py
• src/feature_engineer.py
• src/train_static_only.py
• src/analyze_model.py

Notes

• The current recognizer is optimized for single-hand static poses.
• For full alphabet support, add a temporal model (e.g., LSTM/GRU/Transformer) for motion letters J and Z.
• Performance depends on lighting, camera quality, and consistent hand positioning.