An Evaluation Framework for Multidimensional Multimedia Descriptor Indexing

Automatic multimedia retrieval /equities the use of complex features, which are typically captured by multidimensional descriptors. A basic operation in a multimedia retrieval system is similarity computation, making use of descriptor-dependant metrics. Many data structures have been proposed for managing the representation of multidimensional descriptors, each geared towards efficiency in some set of basic operations. The paper describes a framework for evaluating multidimensional descriptor indexing structures and reports a set of experiments with selected descriptors indexing methods. The extensibility of the framework is illustrated by incorporating a recently-proposed structure, the BitMatrix. Data sets and experiment conditions can be set up so as to provide results that can be used in the choice of appropriate indexing structures for a class of multimedia retrieval applications.