Face Detection with Asymmetric Boosting

Asymmetric boosting, while acknowledged to be§important to state-of-the-art face detection, is§typically based on the trial-and-error practice,§rather than on principled methods. This work solves a§number of issues related to asymmetric boosting and§the use of asymmetric boosting in face detection. It§shows how a proper understanding and use of§asymmetric boosting leads to significant improvements§in the§learning time, the learning capacity, the detection§speed and the detection accuracy of a face detector.§§There are four main contributions in this book: 1) a§new method to learn online an asymmetric boosted§classifier, pioneering a new direction of online§learning a face detector; 2) a new weak classifier§learning method,§significantly reducing the learning time of a§face detector from weeks to just a few hours; 3) a§new and principled method to learn a§face detector cascade, further improving§the learning time and the detection speed of a face§detector; and 4) a theoretical analysis on the§generalization of an asymmetric boosted classifier§via bounds on the true§asymmetric error of the classifier. The work is§concluded with a discussion of future directions for§face detection.