PhotoCheck API

Face detection vs. face recognition

Updated

Face detection finds faces in an image and says where they are. Face recognition works out who a face belongs to by comparing it with others. PhotoCheck does only detection: it returns a box and a confidence for each face, and never computes the face embedding that recognition needs.

What detection gives you

For each face: a bounding box in pixels and a confidence between 0 and 1. That's enough to count people in a photo, require exactly one face in a profile picture, crop a thumbnail, or blur faces before publishing.

PhotoCheck runs BlazeFace, a small, fast detector built for frontal and near-frontal faces, through the open-source Human library. It's tuned with a 0.5 confidence threshold by default; lower it with min_confidence to catch small or side-on faces at the cost of more false positives.

What it deliberately doesn't do

No embeddings, no matching, and no attribute guesses:

  • No identity: it can't say who someone is, or whether two photos show the same person
  • No age, gender, ethnicity or emotion estimates, which are unreliable and easy to misuse
  • No storage: the image and the boxes exist only for the length of the request

Where the confidence goes wrong

Low thresholds find faces in windows, patterns and tree bark. In our own tests, a street of Italian houses produced two 0.3-confidence "faces" that were windows, while real portraits scored around 0.9. That gap is why the default is 0.5.

Check a photo now

Questions

Is face detection personal data under GDPR?

Detection alone doesn't produce biometric identifiers, but the image you send may itself be personal data. PhotoCheck processes it transiently and keeps nothing.

Can it detect masked or side-on faces?

Partly. Lower min_confidence to 0.3 to catch more of them, and expect a few more false positives.

More guides