The Quality vs. Time Trade-off for Approximate Image Descriptor Search

In recent years, content-based image retrieval has become more and more important in many application areas. Similarity retrieval is inherently a very demanding process, in particular when performing exact searches. Therefore, there is an increasing interest in performing approximate searches, where result quality guarantees are traded for reduced query execution time. The goal of approximate retrieval systems should be to obtain the best possible result quality in the minimum amount of time. As a result, typical indexing strategies divide the data set into many data chunks. Minimizing the search time suggests to generate uniformly sized chunks to best overlap I/O costs with CPU costs. Maximizing quality, on the other hand, suggests to strongly limit the intra-chunk dissimilarity of data. The paper addresses the question to what extent guaranteeing the query processing time, using uniform chunk sizes, compromises the quality of the results, and vice versa. Using a large collection of 5 million 24-dimensions local descriptors computed over more than 50 thousand real life images, we show that minimizing the query processing time may in fact lead to better quality of the intermediate results.