camera.ui

English

Deutsch

English

Deutsch

Appearance

Face recognition ​

Face recognition spots faces on your cameras, recognizes people you have enrolled, and groups the ones it doesn't know yet.[1] It builds on object detection, and the camera needs an AI backend in two places under its Plugins tab: Face finds the faces, Face Recognition turns each one into something camera.ui can compare. Without the second one faces are found but never named. If the camera detects objects itself without reporting where they are (many Reolink models), enable Object Assist so faces get a proper crop.

Faces view with known and unknown faces

Enrolling a known face ​

In the Faces view (sections Known Faces and Unknown Faces), Add Face takes a name and a clear photo of the person; Enroll rejects a photo without a face. From then on the person is recognized in new events, with their name as an attribute on detections.

Enroll face dialog

Getting good results ​

  • Camera position. Mount a camera meant for faces at about head height, or with a shallow angle. Looking down steeply it sees forehead and hair, different people start to look alike, and recognition mixes them up.
  • Pictures per person. At least 10 pictures are recommended, from different situations (daylight, night, with a cap). Every further look up to the maximum helps. With fewer the person is often missed, and their dialog says so.
  • Unclear faces get no name. A face seen from the side, from above, or blurred is neither named nor collected under Unknown Faces. The person is recognized in the next picture that shows them clearly.
  • Small faces. A face smaller than about 40 pixels in the camera picture is too small to recognize.
  • One picture is not enough. A person the camera follows gets a name once three pictures of the face agree, and camera.ui keeps checking while they stay in view, so a clear look late in a long visit still counts. When the pictures disagree, the person stays unnamed rather than getting someone else's name. A picture about as close to two enrolled people counts for neither. On a camera whose own AI reports people, faces are only read on its reports, and one picture names them.

A person's pictures ​

A person under Known Faces opens with their training pictures. A mark on a picture says what it does to recognition, tap it for the reason:

  • No clear face. Turned away, from above or blurred. Such a picture makes recognition worse, remove it.
  • Face partly visible. Keep it only while better pictures of that view are missing.
  • Also filed under … The same picture is enrolled for someone else as well. Keep it for the right person only.
  • Same as another picture. One of the two is enough.

Assigning an unknown face warns when it is already filed under someone or only partly visible. Unclear faces never reach Unknown Faces in the first place.

Unknown faces ​

Faces that don't match anyone enrolled collect under Unknown Faces, grouped automatically so the same person's shots cluster together.

  • Assign a group to a person to enroll everyone in it at once, or assign a single face.
  • Discard groups or faces you don't care about, or clear all ungrouped faces.

Correcting a recognition ​

When an event names the wrong person, open that face picture on its recording card, pick Correct person in the dots menu and assign the right name, or Mark as unknown. Only this face changes: other faces in the recording keep their names, even when they carry the same one. The crop is added to that person so recognition gets it right next time. Admins only.

Ignoring a face ​

Ignore a face you never want to see again (a passing stranger, the mail carrier) and camera.ui stops collecting it.

  • Ignore a face or a whole group. Unknown faces and groups each have an ignore button; selection mode ignores several at once.
  • New sightings are dropped. Once someone is ignored, their later shots don't come back as unknown faces.
  • Ignored faces sit in their own list. They collect under Ignored Faces at the bottom of the Unknown Faces section.
  • Restore to undo. Restore on an ignored face lets that person show up again.

Settings ​

In Settings → Recordings, face recognition has:

  • Max Training Images. How many photos to keep per person (5 to 40, default 20). At the limit the picture most like another one of the same person makes way.
  • Max Unknown People. How many unknown groups are kept (10 to 500). Once the limit is hit, the group that was seen longest ago is dropped. Raise it for busy outdoor cameras.

Per camera, Settings → Detection has two face settings:

  • Face confidence. Faces the AI is less sure about than this are ignored for events and automations (0 to 1, default 0.5).
  • Recognize faces. How certain the match against your enrolled people has to be before someone is named: Strict, Balanced (default) or Relaxed. Strict gives fewer wrong names, Relaxed recognizes people more often.

Changing the recognition model ​

The model is a setting of the AI backend (Face Embedding Model), shared by all its cameras. After a change camera.ui embeds the enrolled pictures again by itself. Re-embed faces, bottom left in the Faces view and admins only, starts it by hand. Recognition is patchy until it finishes. Pictures in which the new model finds no clear face are removed, and the dialog names the people this affected.

To try a model, open the AI backend's page and its Face Recognition tab: drop a picture and it shows who camera.ui takes it for and how sure it is.

Rescanning ​

After enrolling someone new, select Rescan existing events to re-check past recordings and tag them with the newly known face.

  1. Face recognition (matching faces against your enrolled people) builds on recording and requires an active camera.ui subscription. Detecting that a face is present is part of the free detection. ↩︎