Active browsing with similarity pyramids

Many approaches to content based retrieval of images have focused on query-by-example methods in which the database is searched for all images similar to an example image presented by the user. However query-by-example techniques tend to quickly converge to a small set of images that may not be of interest. In this paper we present an alternative method for content based search which is based on active browsing using a data structure called a similarity pyramid. The similarity pyramid organizes large databases into a three dimensional pyramid structure which the user can move through. During the browsing process, the user can give feed-back through the selection of a set of desirable images which we call a relevance set. Once selected the relevance set can be used to both prune and reorganize the similarity pyramid to best suit the user's task. A novel cross-validation method is also proposed for effectively performing pruning and reorganization operations.