文档介绍：Real-time Face Detection using FFS boosting method
in Hierarchical Feature Spaces
Hao Ji, Fei Su, Feng Ye, Yuanbo Chen Yujia Zhu
School of Information and munications School of Electronic Engineering
Beijing University of Posts and munications Beijing University of Posts and munications
Beijing 100876, China Beijing 100876, China
e-mail: jihaogood@ e-mail: jas31@
Abstract—AdaBoost based training method has process is really time consuming, and it always
e a state-of-the-art boosting approach in face takes a few weeks to train plete face detector
detection system. In this paper, compared to the cascade. Second, weak classifiers e too weak
naive AdaBoost method, Forward Feature Selection at the later stages of the cascade. Features which
(FFS) method is used in feature selection to reduce were selected in later rounds yielded error rates
the training time by about 50 to 100 times without between and while features selected in early
loss of performance. Furthermore, hierarchical rounds had error rates between and . As the
feature spaces (both local and global) to construct a error of weak classifier approaches 50%, the
detector cascade based on FFS method are adopted, performance of the systems can not be improved by
which still have good discrimination in the later stage adding these weak features.
of boosting pro