GC* -tree: a generic index for perceptual similarity search

Abstract

Similarity search is an intuitive, easy-to-use function, indispensable for MM DBMS. Myriad heuristics were exploited to faithfully represent MM objects for efficient search engine designs. However, the demand to incorporate more representative features for generality incurs great challenge to search speed improvement. A simple linear scan can often outperform many elaborate designs. Particularly, most distance metrics used barely certify perceptual similarity due to contamination of irrelevant features. Alternative non-metric distance functions to selectively choose a dynamic proper subset of holistic features, despite better perceptual accuracy offered, would create non-uniformity against indexing. In this paper, we propose a generic index structure and its supportive search algorithm to attain both accuracy and speed for perceptual similarity search. The idea is to confine the true perceptual distance within the range specified by the upper/lower bound functions we developed. This allows effective pruning and dramatic search time reduction in our solution, which achieves 44% performance gain over linear scan.