Shape matching using a binary search tree structure of weak classifiers

Abstract

In this paper, a novel algorithm for shape matching based on the Hausdorff distance and a binary search tree data structure is proposed. The shapes are stored in a binary search tree that can be traversed according to a Hausdorff-like similarity measure that allows us to make routing decisions at any given internal node. Each node functions as a classifier that can be trained using supervised learning. These node classifiers are very similar to perceptrons, and can be trained by formulating a probabilistic criterion for the expected performance of the classifier, then maximizing that criterion. Methods for node insertion and deletion are also available, so that a tree can be dynamically updated. While offline training is time consuming, all online training and both online and offline testing operations can be performed in O(logn) time. Experimental results on pedestrian detection indicate the efficiency of the proposed method in shape matching.