A model-based discriminative framework for sets of positive vectors classification: Application to object categorization

Classic support vector machines (SVM) kernels (e.g. polynomial, Gaussian) are unable to take advantage of existing problem-specific knowledge. This is especially true when the problem at hand is the classification of sets (or bags) of vectors that may represent textual or visual (e.g. images, videos) data. This article tackles the problem of the classification of sets of positive vectors via SVM. It describes approaches to generate SVM kernels, appropriate for this problem, from generalized inverted Dirichlet (GID) mixtures. We develop several kernels using the fact that the GID belongs to the exponential family of distributions. The promise of such an approach and its advantages are demonstrated via a challenging application that concerns object categorization.