Learning to detect objects of many classes using binary classifiers

Abstract

Viola and Jones [VJ] demonstrate that cascade classification methods can successfully detect objects belonging to a single class, such as faces. Detecting and identifying objects that belong to any of a set of "classes", many class detection, is a much more challenging problem. We show that objects from each class can form a "cluster" in a "classifier space" and illustrate examples of such clusters using images of real world objects. Our detection algorithm uses a "decision tree classifier" (whose internal nodes each correspond to a VJ classifier) to propose a class label for every sub-image W of a test image (or reject it as a negative instance). If this W reaches a leaf of this tree, we then pass W through a subsequent VJ cascade of classifiers, specific to the identified class, to determine whether W is truly an instance of the proposed class. We perform several empirical studies to compare our system for detecting objects of any of M classes, to the obvious approach of running a set of M learned VJ cascade classifiers, one for each class of objects, on the same image. We found that the detection rates are comparable, and our many-class detection system is about as fast as running a single VJ cascade, and scales up well as the number of classes increases. © Springer-Verlag Berlin Heidelberg 2006.