New anchor selection methods for image retrieval

Anchoring is a technique for representing objects by their distances to a few well chosen landmarks, or anchors. Objects are mapped to distance-based feature vectors, which can be used for content-based retrieval, classification, clustering, and relevance feedback of images, audio, and video. The anchoring transformation typically reduces dimensionality and replaces expensive similarity computations in the original domain with simple distance computations in the anchored feature domain, while guaranteeing lack of false dismissals. Anchoring is therefore surprisingly simple, yet effective, and flavors of it have seen application in speech recognition, audio classification, protein homology detection, and shape matching. In this paper, we describe the anchoring technique in some detail and study methods for anchor selection, both from an analytical, as well as empirical, standpoint. Most work to date has largely ignored this problem by fixing the anchors to be the entire set of objects or by using greedy selection from among the set of objects. We generalize previous work by considering anchors from outside of the object space, and by deriving an analytical upper bound on the distance-approximation error of the method.