Efficient Object Detection using Cascades of Nearest Convex Model Classifiers
An object detector must detect and localize each instance of the object class of interest in the image. Current detectors usually adopt a sliding window approach, reducing the problem to one of deciding whether the current detector window contains a valid object instance or background. They typically use binary machine learning discriminants such as SVM for this, and often formulate the problem in terms of a classifier cascade to allow rapid rejection of easy negatives. We argue that in the earlier stages of the cascade, it is more efficient to focus principally on modeling the positive class as this leads to simpler classifiers and faster rejection. We implement this in the form of a short cascade of efficient ``one-class'' nearest convex model classifiers, starting with linear distance-to-affine-hyperplane and interior-of-hypersphere classifiers and finishing with kernelized hypersphere classifiers. We show that our methods have very competitive performance on two face datasets including Labeled Faces in the Wild and state-of-the-art performance on the INRIA Person dataset. As expected, the one-class formulation provides significant reductions in classifier complexity relative to the analogous two class approaches.