audit/vision.md

Vision & Face Recognition Audit

The Face Recognition system is located in backend/Face_Recognition.

Key Findings:

1. Matcher Engine: FaceMatcher handles operations via InsightFace models.
2. Embedding Storage: Enrollments are saved as .npy files. Stored inside faces_db and temp_faces_db.
3. Concurrency: It employs thread locks (threading.Lock()) for updating the embedding dict safely across background refreshes and API calls.
4. Background Refresh: Boot sequence kicks off a thread to refresh stale embeddings without blocking startup.
5. Dynamic Thresholds: Reads defaults via env vars (FACE_MATCH_THRESHOLD) but overrides them based on a PERSON_THRESHOLDS mapping for predefined demo identities (mk, urvi, vidit), ensuring fewer false positives.

Recommendations:

• Transition from file-based .npy lookups to a vector database if the identity count exceeds a few thousand.
• Ensure the background thread handles InsightFace GPU resource contention if any.