A Greedy Approach for Building Classification Cascades

Classification cascade is a well-known technique to reduce classification complexity (recognition time) while attaining high accuracy. While cascades are usually built using ad-hoc procedures, in this paper we introduce a principle way of building cascades using a greedy approach. Given a large pool of classifiers, our approach sequentially builds a near-to-optimal cascade. The approach is fully automated, fast, and scales to large number of classifiers in the pool.