Face recognition identifies or verifies a specific person from the appearance of their face, and it is fundamentally different from face detection. Detection only finds that a face is present; recognition asks who it is. The pipeline detects a face, aligns it, converts it into a numeric "template" (an embedding — numbers, not a stored photo), and matches that template against a gallery. Two modes exist: 1:1 verification (is this the person they claim?) and 1:N identification (who, if anyone, in this watchlist?).
Accuracy must be stated honestly and in context. On cooperative, well-lit, frontal images, top algorithms score very high in NIST testing (around 99%), but real surveillance — angled, distant, poorly lit CCTV — performs far worse, commonly in the 65–85% range, with documented accuracy differences across demographic groups (NISTIR 8280). At scale the false-positive maths bites: a 0.1% false-match rate against a million daily faces is on the order of a thousand wrong hits a day. It is never 100%.
Above all, face recognition is a legal gate, not just a feature. A face template is biometric data under GDPR Article 9 (with a DPIA required under Article 35 and EDPB Guidelines 3/2019); the EU AI Act prohibits most real-time remote biometric identification in public (in force since 2 February 2025); and Illinois BIPA (740 ILCS 14) requires consent and grants a private right of action with statutory damages. Confirm the legal basis before the technical design. This is engineering guidance, not legal advice — confirm specifics with qualified counsel; the model internals belong to the AI for Video Engineering section.
