Two Different Questions
| Face detection | Face recognition | |
|---|---|---|
| Question | Is there a face, and where? | Whose face is this, or who does it look like? |
| Output | Boxes and landmark points | An embedding or a name |
| Needs to know anyone? | No | Yes, a gallery of known faces to compare with |
| Everyday example | The yellow box on your camera screen | Face ID unlocking your phone |
| Privacy risk | Low on its own | Higher, because it links faces to identities |
They run in order. Recognition can't start until detection has found a face and marked where its eyes, nose and mouth are.
How Face Detection Evolved
Viola-Jones (2001)
Paul Viola and Michael Jones built the first detector fast enough for real time on ordinary hardware. It scanned the image for simple light-dark patterns (the eye region is darker than the cheeks; the bridge of the nose is lighter than the eyes) using thousands of tiny tests arranged in a cascade, so most of the image was rejected after a few quick checks. It powered the face boxes on digital cameras for years, but struggled with turned heads and poor light.
Deep learning detectors
From around 2015, neural networks took over. MTCNN (2016) used a cascade of three small networks and also predicted five landmarks. RetinaFace (2019), from the same InsightFace team behind ArcFace, predicts boxes, landmarks and more in a single pass and works on tiny, blurry and partly hidden faces. These detectors are trained and tested on WIDER FACE, a dataset of 32,203 images with 393,703 labelled faces in crowds, at all scales.
Read more: How Does AI Facial Recognition Actually Work? A Plain-English Guide
The Bridge: Landmarks and Alignment
Modern detectors don't just draw a box. They mark landmark points. Ollie uses YuNet, which finds five: both eye centres, the nose tip and both mouth corners.
Those points feed alignment, the step between detection and recognition. The face is rotated, scaled and shifted so the five points land on fixed positions in a 112 by 112 pixel crop. The recognition network was trained on crops aligned exactly this way, so getting this step right matters as much as the network itself. See the maths of face recognition for how the transform is calculated.
How Recognition Works
Recognition takes the aligned crop and turns it into numbers. Ollie's network is a 20-layer convolutional network trained from scratch on about 5.8 million photos of 85,742 people. It outputs 512 numbers, an embedding, and two faces are compared by how close their embeddings are.
For Ollie, recognition doesn't mean naming you. It means comparing your embedding with the stored embeddings of 40,000+ celebrity photos and returning the closest. For the full walk-through, see how face recognition works.
Read more: The Math Behind Your Face: How Machines Read Bone Structure
Get new guides by email
Reading Errors: Which Step Failed?
Knowing the two steps helps you fix a bad result.
"No face found" means detection failed
- The face is turned too far to the side.
- It's too small in the frame, or too blurry.
- It's too dark, or backlit into a silhouette.
- Sunglasses, a mask or hands hide the landmarks.
A face was found but the matches feel off: a recognition-quality problem
- A close selfie distorted your proportions.
- Side lighting changed how your features read.
- A big expression reshaped your eyes and cheeks.
- The wrong face was picked: Ollie uses the largest face, which in a group photo might not be you.
Both are usually fixed with one clear, front-facing photo in soft light. See the best photo for a celebrity match.
Why the Difference Matters for Privacy
Detection alone doesn't identify anyone. It's used to blur faces in street-view photos, focus cameras and count people in a queue. Recognition is the step that links a face to an identity, which is why most laws and debates focus on it.
Ollie keeps recognition narrow on purpose: it compares you only with a fixed list of public figures, never tells anyone who a stranger is, and doesn't store your photo. See privacy and face recognition.