Boosted multi-hypothesis sequential probability ratio test

In many computer vision classification and detection problems, the time to decision is no less important that the error rates. In the binary case, given the false positive and false negative rates, the optimal strategy minimizing the expected decision time is the Wald’s sequential probability ratio test (SPRT). Recently, Sochman and Matas proposed learning SPRT strategies from labeled examples by a boosting technique they dubbed WaldBoost [3]. WaldBoost was shown to outperform state-of-the-art methods in the average evaluation time while keeping comparable error rates. In this paper, we present the extension of WaldBoost to multi-hypothesis classification problems.