AdaBoost and Face Detection ( Version 1 . 0 )

In this report we summarise our experiments with ADABOOST and its application to the problem of face detection. As a starting point, the Viola-Jones face detector ([1]) is reimplemented and some details of the implementation are discussed. Next, modifications of both the training and detection procedure are reported. In the learning phase, tenfold decrease in training time, without noticeable reduction in performance, is achieved via randomisation. Two more modifications of the Viola-Jones algorithm are reported. First, the benefits of illumination normalisation are explored. Improvements both in the speed of training and in detection are achieved at minimum computational costs, since the ’integral image’ technique is employed. Secondly, weak classifiers based on LDA in the space constructed from pairs of Viola-Jones are incorporated into ADABOOST and their performance tested. Finally, work in progress and future directions of our research are discussed.