Face Recognition
Decide whether two faces are the same person
WHAT THIS CAPABILITY MEANS
Takes one or two face images and outputs an identity decision: verification answers whether the two are the same person, while identification asks who in a gallery a face matches. Faces are first aligned and encoded into feature vectors, then compared by distance. Unlike classification the class set is not fixed; the comparison target comes from a dynamic gallery.
How it is done
The pipeline typically detects a face, aligns it by keypoints, and extracts a feature vector with a deep network trained by metric learning to pull the same person together and push others apart. Early systems relied on hand-crafted features; deep metric learning around 2014 pushed accuracy past the practical threshold. Liveness detection, which separates a live person from a photo or a mask, is an indispensable part of any deployment.
Representative products
1Organizations involved
Typical uses
- Access control and ID verification
- Automatic people grouping in photo albums
- Identity confirmation for remote account opening
- Video search and person tracking
How it is evaluated
- 1:1 verification accuracy
- Correct same-or-different decisions
- TAR@FAR
- True accept rate at a fixed false accept rate
- Identification mAP
- Ranking quality when searching a gallery
Limits and hard parts
- Large pose differences, age gaps and harsh lighting push the same person apart in feature space
- Uneven training data produces systematic bias across skin tones and genders
- Photos, replayed video and realistic masks can fool the system, so liveness detection is required, not optional
Concepts behind it
Image Representation
To a machine, a photo is nothing but stacked grids of numbers
Self-Supervised Vision & Contrastive Multimodal Learning
No labels needed: learning to see by working out which images are the same
Model Evaluation & Cross-Validation
Accuracy is the easiest metric to fool you — get evaluation wrong and everything else follows