Fast Evaluation Techniques for Complex Similarity Queries

Complex similarity queries, i.e., multi-feature multi-object queries, are needed to express the information need of a user against a large multimedia repository. Even if a user initially issues a single-object query over one feature, a system with relevance feedback will automatically generate a complex similarity query. Relevance feedback is only useful if response times are interactive. Therefore, this article contributes to the important problem how to evaluate such complex queries efficiently. We describe a new evaluation technique called Generalized VA-File-based Search (GeVAS). It builds on the VA-File [27], supports queries over several feature types, and borrows the idea to search an index structure with several query objects in parallel from Ciaccia et al. [8]. Our main contributions are twofold: 1) we show that GeVAS does not degenerate for queries with many objects or many feature types. 2) We develop a number of variants of GeVAS, tailored to the different distance measures and distancecombining functions, and we show that they yield a significant performance improvement.