How Face Identification Works

In short: the program doesn't decide who's in a photo or video frame. It finds similar faces, shows them to whoever's doing the review, and remembers every decision made. A decision is never lost and never overwrites itself — which matters most when a review gets handed off from one staff member to another, or stretches over months.

How it's built

The machine suggests, the person decides. Face similarity is a hint, not a conclusion. Next to each name you see how close the match is, plus the reference photo it's being compared against. A "confirm all at once" button for a group exists, but it's labeled honestly — not offered as a safe default action.

There's no safety threshold for similarity, and the interface says so directly. Even among matches the system scores as very similar, a meaningful share turn out to be real rejections once a person actually checks — and among weak matches, occasional confirmations still show up. That's why bulk-rejecting a weak group is safe, but bulk-confirming a strong one isn't: the gap between "probably" and "certainly" is smaller than it looks in this task.

Every decision lands in a shared log instead of replacing the previous one. Undo a confirmation, and the undo gets recorded — the original entry doesn't get erased. The result for each face is always recalculated from the full decision history, so you can see not just "what it is now" but "how it got there." For an organization, that means a review can be handed to a different staff member without losing context, and without the risk of someone's earlier work silently getting overwritten.

"Not them" means only "not them." Not "this person doesn't matter to the archive." The photo stays in the queue, waiting for the right name. A rejected name is never suggested again for that photo — not by face similarity, not by a match on the filename.

Any action can be undone. A photo removed by mistake goes back into the queue. Renaming a person doesn't erase the decisions already made about them.

The method's honest limit

Whoever's doing the review usually knows some of the people in the archive personally, and not the rest. For an unfamiliar face, no amount of similarity substitutes for actually knowing who it is — only someone who's actually seen that person can confirm a name. That's why the system has a dedicated "needs a second opinion" status: the card doesn't get lost and doesn't get passed off as identified — it goes to whoever knows the archive better, a specific person, not a guess.

That's not a shortcoming. A tool that pretends it can identify a stranger is lying about it from the start.

What the module doesn't do

• It doesn't search external databases or the internet for people — it works only with what's in the archive and the names a person entered.
• It never confirms anyone on its own, without a person involved.
• It doesn't export every detected face as separate files — it keeps a handful of reference photos per person, not a full gallery of every match found.

Proven in practice

The method and the confirmation interface have been proven on a review covering upward of twenty thousand detected faces and a couple hundred confirmed people — not in demo mode, on a real working archive. The module is already in use wherever people show up on the record: media production and ad agencies, TV studios, TV channels and show production, donated collections at libraries and archives.