A Fast Dual Method for HIK SVM Learning

Histograms are used in almost every aspect of computer vi-4 4 sion, from visual descriptors to image representations. Histogram Inter-5 5 section Kernel (HIK) and SVM classifiers are shown to be very effec-6 6 tive in dealing with histograms. This paper presents three contributions7 7 concerning HIK SVM classification. First, instead of limited to integer8 8 histograms, we present a proof that HIK is a positive definite kernel for9 9 non-negative real-valued feature vectors. This proof reveals some inter-10 10 esting properties of the kernel. Second, we propose ICD, a deterministic11 11 and highly scalable dual space HIK SVM solver. ICD is faster than and12 12 has similar accuracies with general purpose SVM solvers and two recently13 13 proposed stochastic fast HIK SVM training methods. Third, we empir-14 14 ically show that ICD is not sensitive to the C parameter in SVM. ICD15 15 achieves high accuracies using its default parameters in many datasets.16 16 This is a very attractive...

• Type: Conference
• Year: 2010
• Where: ECCV
• Keywords: Computer Vision, ECCV 2010, HIK SVM, HIK SVM Classification, Svm Solvers