Description of interest regions with oriented local self-similarity

Two novel approaches for extracting distinctive invariant features from interest regions are presented in this paper, i.e., Oriented Local Self-Similarities (OLSS,C) and Simplified and Oriented Local Self-Similarities (SOLSS,C) based on Cartesian location grid and gradient orientation for binning, which are the modified versions of the well-known Local Self-Similarities (LSS,LP) feature based on Log-Polar location grid. They combine the powers of well-known approaches, i.e., the SIFT and the LSS (LP), and are achieved by adopting the SIFT algorithm and using the novel LSS and the proposed simplified LSS feature instead of original gradient feature used in SIFT. Furthermore, a new binning strategy for creating feature histogram is proposed where the gradient orientation for binning is calculated from a larger patch in the diagonal direction. The performance of these oriented OLSS (C) and SOLSS (C) descriptors to image matching is studied through extensive experiments on the INRIA Oxford Affine dataset. Empirical results indicate that the proposed OLSS (C) and SOLSS (C) descriptors yield more stable and robust results, significantly outperform the original LSS (LP) descriptor, and also achieve better performance to the SIFT in these experimental evaluations with various geometric and photometric transformations.