Filtering Methods for Similarity-Based Multimedia Retrieval

A common problem in multimedia databases is retrieving the most similar matches to a query object. Finding those matches can be too slow to be practical, especially in domains where comparing multi- media objects involves computationally expensive similarity (or distance) measures. Filter-and-rene retrieval is a framework for addressing this problem: the lter step quickly lters out most database objects, and the rene step identies the best matches among the remaining candidates. This paper describes two ltering methods, that work by constructing ef- cien t approximations of computationally expensive similarity measures. The rst method can be applied to arbitrary domains, and the second method explicitly targets domains where measuring similarity includes an alignment process. The benets of these two ltering methods are illustrated in experiments with databases from dieren t domains, i.e., hand images, gesture videos, and online digit recognition for hand-held devices.