Exploring Relevance Vector Machines for Faster Pedestrian Classification

摘要

While (linear) Support Vector Machines (SVMs) are one of the mainstream choices for pedestrian classification, this work explores the potential benefit of using Relevance Vector Machines (RVMs). Thanks to the sparser representation of RVMs than that of SVMs, it is found that when classifying with a radial-basis function kernel, a tenfold speed-up is obtained with only a slight degradation of the overall discriminative power. However, the training time of RVMs for this problem turns out to be about two orders of magnitude higher than that of SVMs. But, by simply partitioning the training set into subsets and learning several RVMs, we show that the training time of RVMs can be reduced as much as one order of magnitude, with a minor decay in performance, with respect to the single RVM on the full training set. These findings are encouraging to further study RVMs as a promising learning module beyond the current (linear) SVMs.