Tuning Identification

Identification quality is a chain: camera image → detection → matching threshold → clearance event. Tune it in that order — no threshold setting rescues a camera that delivers small, backlit faces, so exhaust Camera Selection & Placement before touching scores.

Thresholds: the false-accept / false-reject trade

Matching thresholds are configured in VPP and express a single trade-off: a higher threshold means fewer false accepts but more travelers who fail to match and fall into identification.no_match; a lower one clears more people at the cost of security margin. Border deployments typically start conservative on denied watchlists (a red alert that is wrong is expensive for the traveler) and tune allowed-watchlist thresholds against observed walk-through scores. Use VPP's identification history to see real score distributions from your site before moving anything.

Liveness and quality gates

Passive liveness (PAD) and ICAO-style quality validation act as gates before matching. Stricter settings protect against spoofing but raise the rate of identification.condition_violation events — faces obscured by caps, masks, or bad angles. Watch that topic's frequency per unit; a spike usually means a physical problem (light, angle, signage) rather than a parameter one.

Enrollment quality matters as much

Match scores are only as good as the enrolled reference: frontal face, good lighting, adequate resolution. If a specific traveler repeatedly scores low everywhere, re-enroll with a better image before blaming thresholds. Watchlist hygiene — removing duplicates and outdated records — measurably improves both accuracy and speed.

Measure, change one thing, measure again

The event stream is your instrument: rates of match vs no_match vs condition_violation per unit, over comparable time windows, before and after each change. Change one parameter at a time and give it enough traffic to be significant. The exact threshold and liveness parameters are documented in the VPP platform documentation.