Re-identification (re-ID)

Same person, no name

Re-ID answers “is this the same person as the one by the till at 4pm”, not “who is this”. The system builds a signature of how a person looks (clothing, build, sometimes the way they move) and matches it against people seen on other cameras or at other times. Nobody gets named; the match connects sightings to each other rather than to a person in the outside world, which is what separates re-ID from facial recognition.

How re-ID systems are built

Most commercial re-ID works on appearance embeddings: a model turns each detected person into a numerical fingerprint and treats close matches across cameras as the same individual. Some systems add gait or height estimation to survive a change of coat. However it's engineered, the output is software that follows particular people around, and that's what gives the technique its regulatory weight.

Svid deliberately does none of this

Svid does no re-identification, no people tracking and no biometrics of any kind. The only AI in the product describes images in plain English, checks conditions on them, and embeds the descriptions for search. You can search footage for “person in a red jacket near the till” and jump to the moments that match; nothing follows that person from camera to camera or builds a profile of them.

Why an operator might prefer that

Tracking and biometric identification of individuals sit in the highest-risk category of processing under UK GDPR and the ICO's guidance on video surveillance, with the heaviest expectations for assessment and justification. An operator whose software never does either has far less to assess and defend. That's a description of the approach, not a compliance certificate; cameras still need signage, a retention policy and a lawful basis, whatever software reads the footage.