Face Learning & Detection

The face detection and registration pipeline running on Whisper-bot handles real-time facial feature landmark extraction, face detection alignment, and vector database embeddings for personalized face recognition.

1. Pipeline Stages

Frame Capture: Real-time video frame retrieval in RGB/BGR formats.
Face Landmarks & Alignment: CNN-based 68 face landmark detection targets nose, eyes, edges, and mouth.
Embeddings Generator: Feeds cropped, aligned face snapshots into a ResNet model (MobileNetV3) to generate a 128-D embedding vector.
Metric Learning: Triplet loss function optimizes distance bounds to match positive profiles and reject negative matches.
Database Storage: Embeddings are mapped to name databases to match faces locally in future frames.

Face Detection & Landmark Pipeline

Face Landmarks & Bounding Box Pipeline