Skip to main content
Version: V12

Understanding Face Recognition

Detection tells you a person is in frame. Face recognition tells you which person. It is the capability that takes a face seen on a live camera and matches it against the people you enrolled in the Object Library, so the moment someone on your watchlist walks past a camera, the system names them, on the video, in the event feed, and in the recording, without an operator having to recognize the face themselves.

Face recognition is the core of the Object Library. This article explains how it works, what makes a recognition reliable, where recognitions show up, and the limits worth knowing.

How Face Recognition Works​

Recognition runs in two stages, enrollment and matching, and they have to agree on what a face "looks like" as numbers.

  1. Enrollment. When you add a person to the Object Library as a Face subject, the AI service finds the face in each reference photo, crops to it, and turns it into a compact numeric signature, an embedding, that captures the face's distinctive geometry. That signature, not the photo itself, is what matching compares against.
  2. Matching. While a camera is live, every face the camera detects is turned into a signature the same way and compared against the signatures of the faces you enrolled. The comparison produces a match score, how close the two signatures are. When the score clears the recognition threshold, the live face is recognized as that enrolled person.

Because enrollment and live matching use the same model to build signatures, a face enrolled from a clear photo can be recognized later from a live frame, even though the two images were never identical.

What Makes a Recognition Reliable​

A recognition is only as good as the reference photos behind it. A few practical things move the needle:

  • Clear, front-facing reference photos. A sharp, well-lit, head-on face produces a strong signature. Small, blurry, heavily angled, or partially covered faces produce weak ones and lead to missed or uncertain matches. This is what the quality score on each enrolled photo is warning you about.
  • More than one photo. Enrolling several photos of the same person, different angles, lighting, with and without glasses, gives matching more to compare against and makes recognition more robust to how the person appears on a live camera.
  • Camera placement. Recognition works best where cameras catch faces reasonably close and head-on, an entrance or a lobby, rather than a high, far, top-down view where faces are small and angled.

Confirming a Match Before Acting​

A single frame can be misleading, a face turns, the light shifts, someone passes behind a pillar. Rather than react to one lucky frame, the system watches a tracked face across several frames and confirms the identity before it commits to it. It also re-checks a recognized face periodically while the person stays in view, and clears the identity when they leave. The result you see is a steady, confirmed name, not a label that flickers on and off frame by frame.

Where Recognitions Show Up​

Once a face is recognized, the recognition flows into the same places you already watch:

  • Live player. The recognized person's name appears on the bounding box drawn around their face, so you see who it is as you watch, not just that there's a person.
  • Event feed. The recognition is listed as its own entry, separate from plain detections, so you can scan who was recognized and when.
  • Recordings. A recognition can trigger a recording, so the moment a watchlisted person appears is captured as reviewable evidence. Whether recognitions drive recording is controlled in your surveillance recording settings, independently of plain object detections.
  • The subject's activity. Each recognition is logged against the enrolled person in the Object Library, building the record of where and when they were seen.

Privacy and Isolation​

  • Your gallery is yours. Faces enrolled in your portal are matched only against your portal's cameras. They are never compared against another tenant's cameras, even on shared infrastructure.
  • Signatures, not just photos. Recognition compares mathematical signatures derived from faces. Treat enrolled photos and recognition data as sensitive personal data and govern them under your organization's privacy and retention policies.

Key Considerations​

  • Face is the recognition type available today. The Object Library lets you organize subjects as Face, Person, Vehicle, or Object, but live matching currently runs for faces. Recognition for the other types is planned; the feature is built on a common framework so those models plug into the same enrollment and matching flow as they ship.
  • Recognition depends on detection. A face is only recognized on a camera that is detecting faces in the first place. Make sure face detection is enabled on the cameras where you want recognition. See How Live AI Detection Works.
  • Changing the underlying model means re-enrolling. If the recognition model is upgraded, signatures from the old model are not comparable with the new one, and affected subjects must be re-enrolled from their photos. The system guards against mixing signatures from different model versions.
  • It assists operators; it does not replace them. Treat a recognition as a strong, scored signal to review, not an automatic identification.

See Also​

Did you find this article helpful?