Generalized One-Class Learning Using Pairs of Complementary Classifiers

In this paper, we present novel objectives for one-class learning, which we collectively refer to as Generalized One-class Discriminative Subspaces (GODS). Our key idea is to learn a pair of complementary classifiers to flexibly bound the one-class data distribution, where the data belongs to the positive half-space of one of the classifiers in the complementary pair and to the negative half-space of the other. To avoid redundancy while allowing non-linearity in the classifier decision surfaces, we design each classifier as an orthonormal frame and learn these frames via jointly optimizing for two objectives, namely: i) to minimize the distance between the two frames, and ii) to maximize the margin between the frames and the data. The learned frames will thus characterize a piecewise linear decision surface allowing for efficient inference, while our objectives seek to bound the data within a minimal volume that maximizes the decision margin, thereby robustly capturing the data distribution. We explore several variants of our formulation under different constraints on the constituent classifiers, including kernelized feature maps. We provide experiments on several applications in computer vision, including anomaly detection in video sequences, human poses, and activities, as well as on five UCI datasets, demonstrating state-of-the-art results.