Class-specific representation based distance metric learning for image set classification

Abstract

Image set classification, which compares the similarity between image sets with variable quantity, quality and unordered heterogeneous images, has drawn increased research attention in recent years. Although many effective image set classification algorithms have been developed, they struggle to overcome issues such as intra-set diversity, inter-set similarity, and input data that are not linearly separable. In this paper, we propose a class-specific representation based distance metric learning (CSRbDML) framework to improve the classification performance. Specifically, CSRbDML aims to learn an inter-set distance metric on a kernel space such that the distance between truly matching sets is smaller than that between incorrectly matching sets. Furthermore, we also propose a novel and powerful image set classifier based on the learned distance metric. Extensive experiments on several well-known benchmark datasets demonstrate the effectiveness of the proposed methods compared with the existing image set classification algorithms.