Increasing retrieval efficiency by index tree adaptation

Image databases often operate in a query-by-example mode where images are retrieved according to feature (dis-) similarity to an example image. Retrieval efficiency is increased by using indexing trees such as kd-trees, quad trees, or R*-trees. However such trees are usually constructed without reference to the similarity measure, and in practice, their performance degrades when the threshold on the similarity value increases beyond zero. This phenomenon is analyzed in this paper with a probabilistic model and an expression is obtained for the average computation in the tree. Based on this analysis, a greedy algorithm is proposed which adapts the tree by eliminating inefficient nodes. The greedy algorithm is based on a ''Markovian'' property of indexing trees. The algorithm is iterative and is guaranteed to improve the performance of the tree with every iteration. Experimental evaluation of the performance of adapted trees for randomly distributed data is reported. The experiments indicate that the performance of the tree improves significantly after adaptation.