Efficient Indexing Of High Dimensional Normalized Histograms

This paper addresses the problem of indexing high dimensional normalized histogram data, i.e., D-dimensional feature vectors H where Sigma(i=1)(D) H-i = 1. These are often used as representations for multimedia objects in order to facilitate similarity query processing. By analyzing properties that are induced by the above constraint and that do not hold in general multi-dimensional spaces we design a new split policy. We show that the performance of similarity queries for normalized histogram data can be significantly improved by exploiting such properties within a simple indexing framework. We are able to process nearest-neighbor queries up to 10 times faster than the SR-tree and 3 times faster than the A-tree.