Learning Policies for Efficiently Identifying Objects of Many Classes

Viola and Jones (VJ) cascade classification methods have proven to be very successful in detecting objects belonging to a single class - e.g., faces. This paper addresses the more challenging 'many class detection' problem: detecting and identifying objects that belong to any of a set of classes. We use a set of learned weights (corresponding to the parameters of a set of binary linear separators) to identify these objects. We show that objects within many real-world classes tend to form clusters in this induced 'classifier space'. As the results of a sequence of classifiers can suggest a possible label for each object, we formulate this task as a Markov decision process. Our system first uses a 'decision tree classifier' (i.e., a policy produced using dynamic programming) to specify when to apply which classifier to produce a possible class label for each sub-image W of a test image. It then uses a cascade of classifiers, specific to each 'leaf' in this tree, to confirm that W is an instance of the proposed class. We present empirical evidence to verify that our ideas work effectively: showing that our system is essentially as accurate as running a set of cascade classifiers, but is much faster than that approach