Efficient image representation for object recognition via pivots selection

摘要

Patch-level features are essential for achieving good performance in computer vision tasks. Besides well-known pre-defined patch-level descriptors such as scalein-variant feature transform (SIFT) and histogram of oriented gradient (HOG), the kernel descriptor (KD) method offers a new way to "grow-up" features from a match-kernel defined over image patch pairs using kernel principal component analysis (KPCA) and yields impressive results. In this paper, we present efficient kernel descriptor (EKD) and efficient hierarchical kernel descriptor (EHKD), which are built upon incomplete Cholesky decomposition. EKD automatically selects a small number of pivot features for generating patch-level features to achieve better computational efficiency. EHKD recursively applies EKD to form image-level features layer-by-layer. Perhaps due to parsimony, we find surprisingly that the EKD and EHKD approaches achieved competitive results on several public datasets compared with other state-of-the-art methods, at an improved efficiency over KD.