Face Recognition Search

Face detection, face recognition and face search

The terms describe successive steps:

Step	What it does	Result
Face detection	Finds the faces in an image	The position of each face, with no identity
Face embedding	Converts each detected face into a vector that describes its features	One vector per face
Face recognition	Compares a face's vector with the vectors of faces already named	A match to a known person, or none
Face search	Filters the library to the images of one recognised person	Every photo of that person

Detection alone does not identify anyone. Identity is added when a person gives a face a name. From then on, recognition can propose that name for faces that look like it.

What teams use it for

• Press requests. A journalist asks for approved photos of the CEO, and communications needs every current portrait, not only the ones in the press folder.
• Events. A photographer delivers 2,000 photos from a conference, and each speaker gets their own selection.
• Leavers and withdrawn consent. When an employee leaves or a model withdraws consent, the team finds every image they appear in and removes it from shares, portals and sales decks.
• Release checks. Before a campaign uses a photo, the team checks that every recognisable person in it has a model release.
• Avoiding reshoots. Before booking a photographer, check whether the library already holds a usable photo of the person.

Face recognition and data protection

A face template used to identify a person is biometric data. Under the GDPR, biometric data processed to uniquely identify a person is a special category of personal data (Article 9), and processing it needs one of the conditions in Article 9(2), most often the person's explicit consent. A data protection impact assessment (Article 35) is often required as well.

Practices that keep the processing under control:

• Names are given to faces by a person, deliberately. The system does not create identities by grouping unknown faces on its own.
• Consent and model releases are recorded with the images, together with the date consent was withdrawn.
• Only named roles can identify faces and search by person.
• Face data stays in your own infrastructure, or with a processor covered by a data processing agreement.

Face recognition search in Freedam

• Faces are detected automatically in images after upload, by a face detection service that runs inside Freedam's own infrastructure rather than at an outside AI provider.
• Detected faces stay unidentified until someone with the Edit face identities permission names them. Freedam does not create people on its own.
• After a face is named, Freedam finds unidentified faces that look like the same person and offers to name them all in one step.
• New faces that closely match a named person are named automatically. Administrators set the recognition threshold, or turn automatic matching off.
• Person search in the asset library filters the gallery to one person's photos and combines with other filters such as a collection or a date. Named people are also found by keyword search.
• Face recognition is available on the plans that include it.

Details: Faces and people and Finding photos of a person in the Freedam docs.

Frequently asked questions

What is the difference between face detection and face recognition?

Face detection finds where faces are in an image. Face recognition decides whose face it is, by comparing it with faces that have already been named. A system can detect faces without recognising anyone.

Is face recognition legal under the GDPR?

It can be, with a lawful basis and one of the Article 9(2) conditions, usually explicit consent from the people concerned. Check with your data protection officer before you enable it, and record consent with the images.

How accurate is face recognition?

Current models are accurate on clear photos where most of the face is visible. Accuracy drops with profile views, low resolution, sunglasses and heavy shadows, and for people who look alike, such as relatives. Check a few matches after naming someone, and raise the matching threshold if different people are confused.