FALCON: Feedback Adaptive Loop for Content-Based Retrieval

Abstract

Several methods currently exist that can perform relatively simple queries driven by relevance feedback on large multimedia databases. However, all these methods require that objects be represented as vectors within feature spaces, and none have yet been designed which handle queries consisting of multiple, arbitrarily shaped regions. We propose a novel method that is designed to handle disjunctive queries within metric spaces. The method relies on users specifying examples of the desired items within the database; retrieving a ranked list of results; and providing feedback on the top-ranked items. We propose a novel way to estimate the dissimilarity of an object to a set of desirable objects. This method differs from many existing feedback methods that rely solely on the Euclidean or Mahalanobis metrics, as it can (a) learn concave shapes within the feature space; (b) learn disjunctive queries, and (c) work within arbitrary metric spaces unsuitable for other methods. This paper demonstrates how the proposed method can be used to "learn" several different types of queries, and shows that our method achieves good precision versus recall within a few iterations on a variety of data sets and queries. Moreover, we show how to use state-of-the-art database indexing methods to achieve up to a 200% improvement in speed over sequential scanning when retrieving up to 20 objects.

More information

Identifiers

Catalogue record

Export record