Atlas record: https://www.aitimeline.dev/en/events/evt-0201/
Research · December 2001
Viola-Jones face detection
The method detected faces in real time on an ordinary processor, discarding most of the image through a cascade of increasingly accurate checks.
Why it matters
Computer vision produced a result cheap enough to put in a device, and cameras started finding faces.
Three ideas together: the integral image, giving the sum of any rectangle in four lookups; feature selection by boosting; and a cascade whose early stages are cheap and reject almost everything. The method sat in compact cameras, phones and OpenCV for years. It stayed the standard until convolutional networks learned to find features themselves.
Event record
- Event date
- December 2001
- Timeline date
- Event date
- Verification
- Sources gathered automatically · September 17, 2026
- Lines
- ID
- evt-0201
Presented at CVPR in December 2001.
Related events
Earlier
- Related Mobileye is founded 1999links to this record
The same bet: real-time recognition on cheap hardware rather than on a workstation.
- Enables A neural network finds faces in photographs 1995links to this record
Viola and Jones measure their detector on the MIT+CMU test set of 130 images and 507 faces from this paper and compare it with the Rowley-Baluja-Kanade detector, which they call the fastest of its time.
- Builds on AdaBoost 1995
The cascade of classifiers is trained by boosting itself.
Later
- Contrasts Gradient histograms find pedestrians 2005links to this record
The paper measures itself explicitly against the Haar-feature AdaBoost detector and shows at least a tenfold reduction in false positives on its own set.
- Builds on Faces from the news as a recognition benchmark 2007links to this record
The only filter the photographs passed through was the Viola-Jones face detector.
- Related AlexNet at ImageNet 2012
The features here are hand-designed; convolutional networks later learn them.
- Builds on Face recognition accuracy by sex, race and age 2012links to this record
The atlas holds the arrival of a face detector that worked, but no measurement of whom it worked worse on; this paper measures that across commercial systems descended from that line.