Viola-Jones made face detection fast by rejecting easy negatives first

Face detection once meant scanning many possible windows and running too much computation on every patch. Viola and Jones reframed the problem as a cascade: most windows are background, so make rejecting them nearly free.

Their CVPR paper names three contributions: the integral image for rapid feature computation, AdaBoost to select a small number of critical visual features, and a cascade that spends more computation only on promising object-like regions.

For face detection, the paper reports rates comparable to the best previous systems and says the detector ran at 15 frames per second without image differencing or skin-color detection.

Reject blank image regions almost free

The integral image made many rectangular features cheap enough to test repeatedly.

AdaBoost turned a huge feature pool into a compact detector.

The cascade matched the data distribution: most windows are easy negatives.

Speed can come from ordering decisions, not only from faster features. Put the cheapest high-recall filter first and reserve detail for the few candidates that survive.

The Viola-Jones framework became a landmark in real-time computer vision and face detection. Later deep learning systems surpassed it, but the cascade idea remains a clean example of spending computation only where uncertainty survives.