A Fast Dual Method for HIK SVM Learning

Abstract

Histograms are used in almost every aspect of computer vision, from visual descriptors to image representations. Histogram Intersection Kernel (HIK) and SVM classifiers are shown to be very effective in dealing with histograms. This paper presents three contributions concerning HIK SVM classification. First, instead of limited to integer histograms, we present a proof that HIK is a positive definite kernel for non-negative real-valued feature vectors. This proof reveals some interesting properties of the kernel. Second, we propose ICD, a deterministic and highly scalable dual space HIK SVM solver. ICD is faster than and has similar accuracies with general purpose SVM solvers and two recently proposed stochastic fast HIK SVM training methods. Third, we empirically show that ICD is not sensitive to the C parameter in SVM. ICD achieves high accuracies using its default parameters in many datasets. This is a very attractive property because many vision problems are too large to choose SVM parameters using cross-validation.