Efficient kernel descriptor for image categorization via pivots selection

Patch-level features are essential for achieving good performance in compute vision tasks. Besides well-known predefined patch-level descriptors such as SIFT and HOG, the kernel descriptor (KD) method [1] offers a new way to ‘grow up’ features from a match kernel defined over image patch pairs using kernel principal component analysis (KPCA). However, under this technical construction, all joint basis vectors are involved in the kernel descriptor computation, which is both expensive and not necessary. To address this problem, we present efficient kernel descriptor (EKD), which is built upon incomplete Cholesky decomposition. EKD automatically selects a small number of pivot features to achieve better computational efficiency. Perhaps due to parsimony, we find surprisingly that despite efficiency, the EKD approach achieved superior image/scene categorization performance than the original kernel descriptor approach.