Benchmarking access structures for the similarity retrieval of high-dimensional multimedia data

In multimedia databases it is usual to map objects into feature vectors in high dimensional spaces. In order to speed query processing, access structures, or indices, are required. Classical spatial access structures such as the R*-tree are bound to fail when the space dimensional is not low. Fortunately, several access structures for high dimensional spaces, e.g., the SS-tree, SR-tree and M-tree have been proposed. However, each of those structures have been benchmarked in a rather ad-hoc manner. The paper benchmarks and compares all the above structures using a real dataset of 40000 high dimensional objects. All structures have been implemented on top of the GiST infrastructure to minimize the risk of implementation bias. Even though no structure can be claimed to be the undisputed winner, we have found that the SR-tree presents the best overall results