Timeline at a Glance
| Year | Milestone |
|---|---|
| 1964 to 1966 | Woody Bledsoe, Helen Chan Wolf and Charles Bisson match faces using hand-entered landmarks |
| 1973 | Takeo Kanade builds the first fully automatic face recognition system |
| 1987 | Lawrence Sirovich and Michael Kirby show faces can be compressed with principal component analysis |
| 1991 | Matthew Turk and Alex Pentland publish eigenfaces |
| 1993 to 1997 | The US FERET programme creates the first large face database and evaluations |
| 2001 | Viola-Jones enables real-time face detection; face recognition is used at the Super Bowl in Tampa |
| 2007 | Labeled Faces in the Wild (LFW) benchmark released |
| 2014 | Facebook's DeepFace reaches 97.35% on LFW |
| 2015 | Google's FaceNet reaches 99.63% |
| 2017 | Apple launches Face ID |
| 2019 | ArcFace reaches 99.83%; NIST publishes its major demographic study |
| 2020 | IBM, Amazon and Microsoft pause or stop police sales |
| 2024 | The EU AI Act restricts real-time public face recognition |
The 1960s and 70s: Measuring by Hand
Woody Bledsoe, working with Helen Chan Wolf and Charles Bisson at Panoramic Research in Palo Alto, built the first face recognition system in the mid-1960s. An operator used a RAND tablet, an early graphics tablet, to mark points like the centres of the eyes, the corners of the mouth and the tip of the nose on each photo. The computer then compared the distances between those points. Much of the funding came from US intelligence, and much of the work stayed unpublished for years.
The system couldn't find faces on its own and struggled with head turns, but the core idea, reducing a face to measurements, still underlies modern systems. In 1973, Takeo Kanade at Kyoto University built the first system that found facial features and measured them automatically.
Read more: How Does AI Facial Recognition Actually Work? A Plain-English Guide
The 1990s: Eigenfaces
In 1991, Matthew Turk and Alex Pentland at MIT published "Eigenfaces for Recognition", building on Sirovich and Kirby's work. Instead of choosing which measurements to take, they let the data decide. Principal component analysis found the main directions in which a set of face images varied, which looked like ghostly face patterns, the "eigenfaces". Any face could be described as a mix of them, a point in a "face space".
It was a turning point: the first successful approach that learned what mattered from data rather than from a human's list. Fisherfaces (1997) improved on it by focusing on differences between people rather than all variation. The US government's FERET programme, running from 1993 to 1997, gave researchers a shared database and test, which made progress measurable.
The 2000s: Real Time and Real Use
In 2001, Paul Viola and Michael Jones published a face detector fast enough to run in real time on ordinary computers. It's why digital cameras could suddenly draw boxes around faces. The same year, police scanned crowds at the Super Bowl in Tampa, prompting one of the first big public debates about face surveillance.
Recognition improved with hand-designed features such as Local Binary Patterns (2004), which described small patches of texture in a way less sensitive to lighting. In 2007, the University of Massachusetts released Labeled Faces in the Wild, a benchmark of 13,233 news photos of 5,749 people. Unlike earlier tests, it used real, messy photos, and it became the yardstick for the next decade.
Read more: ArcFace: The Loss Function That Makes Modern Face Recognition So Accurate
Get new guides by email
2014 Onward: Deep Learning
Deep learning changed everything in about two years. In 2014, Facebook's DeepFace scored 97.35% on LFW, close to human performance on the same cropped faces. The Chinese University of Hong Kong's DeepID models pushed past it within months. In 2015, Google's FaceNet reached 99.63% using triplet loss and 200 million training images.
Then came margin-based losses: SphereFace (2017), CosFace (2018) and ArcFace (2019), which reached 99.83% on LFW. These trained like simple classifiers but produced very clean face embeddings. See ArcFace explained. Face recognition moved into phones (Face ID, 2017), airports and police systems.
Ollie's own model belongs to this generation: a 20-layer SphereFace-style network trained from scratch with CosFace loss on 5.8 million face photos, scoring 99.5% on LFW. See how Ollie works.
Where the Debate Is Now
Accuracy on easy benchmarks is essentially solved. The open questions are about fairness and use. NIST's 2019 study showed large demographic differences in many algorithms. Wrongful arrests made headlines. In 2020, IBM left the face recognition business, Amazon paused police sales and Microsoft said it wouldn't sell to US police without a federal law. Cities banned police use, and the EU wrote limits into law.
Sixty years after Bledsoe marked faces by hand, the technology works. The history now being written is about where it should and shouldn't be used. For more, see privacy and face recognition.