Matching takes the probe embedding, the one just computed from the new face, and compares it against stored embeddings for known people. The closest stored embedding gives a similarity score. That score is then checked against a threshold: above it, the system declares a match; below it, no match. The threshold is a trade-off, since raising it reduces false accepts but increases false rejects. The same machinery answers two different questions. Verification is 1:1, asking whether this face is the person it claims to be. Identification is 1:N, asking which of many enrolled people this face belongs to, or whether it belongs to none of them.
How AI Recognizes a Face: From Detection to Identity Match