Acceleration Technique for Boosting Classification and Its Application to Face Detection

Abstract

We propose an acceleration technique for boosting classification without any loss of classification accuracy and apply it to a face detection task. In classification task, much effort has been spent on improving the classification accuracy and the computational cost of training. In addition to them, the computational cost of classification itself can be critical in several applications including face detection. In face detection, a celebrating work by Viola and Jones (2001) developed a significantly fast face detector achieving a competitive accuracy with all preceding face detectors. In their algorithm, the cascade structure of boosting classifier plays an important role. In this paper, we propose an acceleration technique for boosting classifier. The key idea of our proposal is the fact that one can determine the sign of discriminant function before all weak learners are evaluated in general. An advantage is that our algorithm has no loss in classification accuracy. Another advantage is that our proposal is a unsupervised learning so that it can treat a covariate shift situation. We also apply our proposal to each cascaded boosting classifier in Viola and Jones type face detector. As a result, our proposal succeeds in reducing the classification cost by 20%.