Glossary · Technology
Face detection
Face detection is the step that finds where faces are in an image and returns a bounding box and landmark points for each one. It answers only whether and where a face is present, not whose it is. Every face search, verification system and camera autofocus starts with it, and its failures (small, turned or occluded faces) are inherited by everything downstream.
Updated November 6, 2026 · 2 min read
What does a detector output?
For each face: a rectangle, a confidence value, and usually five or more landmarks (eyes, nose tip, mouth corners). The landmarks drive alignment, where the face is rotated and scaled to a standard pose before a face embedding is computed. Detection quality therefore sets a ceiling on everything that follows.
How is it different from face recognition?
Detection says "there is a face here". Face recognition says "this face is similar to that one". A phone camera that draws a square around faces detects; a photo app that groups pictures by person recognises. Privacy law treats the two differently: detection alone usually does not create biometric data about an identifiable person, while recognition does.
Which methods are used?
| Method | Era | Notes |
|---|---|---|
| Viola-Jones cascades | 2001 onward | Fast, frontal faces only, still used in cameras |
| HOG plus SVM | 2005 onward | Better on varied lighting, weak on profiles |
| MTCNN and other CNN detectors | 2016 onward | Joint detection and landmarks, handles pose and scale |
| RetinaFace, SCRFD and transformer detectors | 2019 onward | Dense landmarks, small faces, used in modern search engines |
Where does it go wrong?
- Faces under about 40 to 60 pixels across are often missed or produce poor landmarks, which is why a face search asks for a clear, large face.
- Strong profiles, masks, sunglasses and heavy motion blur lower confidence or drop the face.
- Group photos produce several boxes; a search engine must pick one or ask the user to.
- Detectors trained on unbalanced data perform unevenly across skin tones and ages, a problem documented in accuracy, bias and testing.
Where does it come up in face search?
Twice. At index time, the crawler runs a detector over every image it fetches, and only detected faces enter the index. At query time, the uploaded photo is detected, aligned and embedded. A "no face found" message at upload is the detector declining the photo.
Questions
Does face detection identify me?
No. It locates a face in the frame. Identification requires comparing an embedding of that face with others.
Why did an engine say no face was detected?
The face was probably too small, turned away, occluded or blurred. A larger, front-facing, well-lit photo usually fixes it.
Sources
- 1.Zhang et al., Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Networks (MTCNN) (checked April 11, 2016)
- 2.Deng et al., RetinaFace: Single-stage Dense Face Localisation in the Wild (checked May 2, 2019)
Entry reviewed November 6, 2026. Found an error? Tell us.
Related entries
- Term Face recognition Face recognition defined: comparing faces to decide whether they match, the difference between verification, identification and search, and how it is regulated.
- Term Face embedding Face embedding defined: the numeric vector a network computes from a face, why same-person vectors sit close, how it is compared, and why it is biometric data.
- Concept Face detection and alignment What face detection and alignment do before a face is compared, how detectors work, why small or turned faces fail, and the photo properties that matter.
- Concept Accuracy, bias and testing How face recognition accuracy is measured, what NIST FRTE reports found on error rates and demographic differences, and why marketing claims cannot be checked.
Cite this entry
FaceSearch Wiki editors. "Face detection". FaceSearch Wiki, updated November 6, 2026. https://facesearch.wiki/glossary/face-detection/