Item

Abstract

Recent developments in computer vision have shown that local features can provide efficient representations suitable for robust object recognition. Support Vector Machines have been established as powerful learning algorithms with good generalization capabilities. In this paper, we combine these two approaches and propose a general kernel method for recognition with local features. We show that the proposed kernel satisfies the Mercer condition and that this type of kernel is suitable for many standard local feature techniques in computer vision. Large-scale recognition results are presented on three different databases, which demonstrate that SVMs using the proposed kernel perform better than standard Nearest-Neighbor techniques on local features. In addition, experiments on noisy and occluded images show that local feature representations significantly outperform global approaches.