How barathi-1993/indian-sign-language-recognition works
This repository implements isolated Indian Sign Language gesture recognition from webcam input. In the collection workflow, frames are processed into 1662-dimensional MediaPipe Holistic landmark vectors and saved as sequences; preprocessing validates sequences, encodes labels, and splits datasets. Training builds and fits the MOPGRU classifier, while evaluation reports metrics. Separately, real-time inference buffers webcam landmarks, classifies full sequences, and displays thresholded labels. The README documents these core stages; sampled sources establish the concrete calls shown. The data and models directories exist, but their runtime contents were not sampled.
Main entry points: realtime_inference.py Live Recognition · collect_keypoints.py Keypoint Collection · preprocess_keypoints.py Sequence Preparation · train.py Training and Reports
Components 15
Capture and Features
- Keypoint Collection
src/collect_keypoints.py - Holistic Features
src/mediapipe_utils.py
Dataset Preparation
- Collected Sequences
data - Sequence Preparation
src/preprocess_keypoints.py - Encoded Splits
data - Label Encoder
models
Model and Evaluation
- Training and Reports
src/train.py - MOPGRU Model
src/mopgru_cell.py - Saved Model
models - Model Evaluation
src/evaluate.py - Evaluation Results
Real-Time Inference
- Live Recognition
src/realtime_inference.py - Prediction Display
src/realtime_inference.py
Other components
- Operator
- Webcam
Connections 22
- Operator → Keypoint Collection: starts capture
README.md - Webcam → Keypoint Collection: provides frames
src/collect_keypoints.py - Keypoint Collection → Holistic Features: calls detection and extraction
src/collect_keypoints.py - Keypoint Collection → Collected Sequences: writes sequences
src/collect_keypoints.py - Operator → Sequence Preparation: starts preparation
README.md - Collected Sequences → Sequence Preparation: reads sequences
src/preprocess_keypoints.py - Sequence Preparation → Encoded Splits: writes splits
src/preprocess_keypoints.py - Sequence Preparation → Label Encoder: saves encoder
src/preprocess_keypoints.py - Operator → Training and Reports: starts training
README.md - Encoded Splits → Training and Reports: loads splits
src/train.py - Training and Reports → MOPGRU Model: builds classifier
src/train.py - Training and Reports → Saved Model: saves checkpoint
src/train.py - Operator → Model Evaluation: starts evaluation
README.md - Encoded Splits → Model Evaluation: loads test split
src/evaluate.py - Saved Model → Model Evaluation: loads model
src/evaluate.py - Model Evaluation → Evaluation Results: writes metrics
src/evaluate.py - Operator → Live Recognition: starts inference
README.md - Webcam → Live Recognition: provides frames
src/realtime_inference.py - Live Recognition → Holistic Features: calls detection and extraction
src/realtime_inference.py - Saved Model → Live Recognition: loads model
src/realtime_inference.py - Label Encoder → Live Recognition: loads labels
src/realtime_inference.py - Live Recognition → Prediction Display: renders predictions
src/realtime_inference.py