What is face recognition?

How it works

Detection finds where faces are in an image and draws a box around each one. Alignment rotates and scales each face so the eyes and mouth sit in standard positions. Embedding then turns the aligned face into a vector, a list of numbers trained so that photos of the same person land close together and different people land far apart.

With those vectors, a system can do three different jobs. Verification asks whether two faces are the same person (one-to-one, like unlocking a phone). Identification asks which known person a face belongs to (one-to-many). Clustering groups unknown faces that look alike, which is how photo apps build a “people” view before anyone has typed a name.

An example

A family has 20 years of home videos. Clustering groups every appearance of the same face, across ages and haircuts, into one pile. Someone names the pile “Grandma,” and now every clip where she appears can be listed and opened at the right moment, including ones nobody remembered she was in.

Notice what the system did and did not do. It never knew who Grandma was; it only knew that a set of faces looked alike. The name came from a person. That split, machine-made groups and human-given names, is how most face tools for personal libraries work, and it is why a wrong merge is easy to fix: you move a face out of the group rather than retraining anything.

Limits and responsibilities

Accuracy drops with side angles, poor lighting, low resolution, motion blur, sunglasses and masks, and across large age gaps such as childhood to adulthood. Look-alike relatives and twins get merged. Published evaluations have also found error rates that vary across demographic groups, so matches should be treated as suggestions to confirm, not facts.

Faces are biometric data. Several laws treat them as sensitive, depending on where you live, and people reasonably expect a say in being identified. Get consent before face-indexing recordings of people outside your household, keep the data local where you can, and prefer tools that let you switch it off and delete it.

In MediaFind

People & Faces is a Pro feature that runs entirely on your device; nothing about faces leaves your machine. Click a face to jump to every appearance of that person, and famous public figures are named automatically. You can switch face indexing off or delete all stored face and people data with a single click, and free libraries are never face-indexed. See how to find a person in your videos and family archives.

Frequently asked questions

What is the difference between face detection and face recognition?

Detection only finds that there is a face and where it is. Recognition goes further and works out whose face it is, or whether it matches another one.

Is face recognition legal to use on my own videos?

Rules vary by country and state. Using it privately on your own family footage is generally treated differently from identifying strangers, but check local law and ask for consent where others are involved.

Why does face recognition split one person into two groups?

Big changes in age, angle, lighting or glasses can push two photos of the same person far apart. Merging the groups by hand usually fixes it.