A binary decision tree implementation of a boosted strong classifier

Viola and Jones [1] proposed the influential rapid object detection algorithm. They used AdaBoost to select from a large pool a set of simple features and constructed a strong classifier of the form {Sigma(j) alpha(j) h(j)(x) >= theta} where each h(j)(x) is a binary weak classifier based on a simple feature. In this paper, we construct, using statistical detection theory, a binary decision tree from the strong classifier of the above form. Each node of the decision tree is just a weak classifier and the knowledge of the coefficients alpha(j) is no longer needed. Also, the binary tree has a lot of early exits. As a result, we achieve an automatic speedup that always makes the rapid Viola and Jones algorithm rapider.