Novel similarity measures for differential invariant descriptors for generic object retrieval

Local feature matching is an essential component of many image and object retrieval algorithms. Euclidean and Mahalanobis distances are mostly used in order to quantify the similarity of two stipulated feature vectors. The Euclidean distance is inappropriate in the typical case where the components of the feature vector are incommensurable entities, and indeed yields unsatisfactory results in practice. The Mahalanobis distance performs better, but is less generic in the sense that it requires specific training data. In this paper we consider two alternative ways to construct generic distance measures for image and object retrieval, which do not suffer from any of these shortcomings. The first approach aims at obtaining a (image independent) covariance matrix for a Mahalonobis-like distance function without explicit training, and is applicable to feature vectors consisting of partial image derivatives. In the second approach a stability based similarity measure (SBSM) is introduced for feature vectors that are composed of arbitrary algebraic combinations of image derivatives, and likewise requires no explicit training. The strength and novelty of SBSM lies in the fact that the associated covariance matrix exploits local image structure. A performance analysis shows that feature matching based on SBSM outperforms algorithms based on Euclidean and Mahalanobis distances.