Semiautomatic benchmarking of feature vectors for multimedia retrieval

Abstract

Modern Digital Library applications store and process massive amounts of information. Usually, this data is not limited to raw textual or numeric data - typical applications also deal with multimedia data such as images, audio, video, or 3D geometric models. For providing effective retrieval functionality, appropriate meta data descriptors that allow calculation of similarity scores between data instances are requires. Feature vectors are a generic way for describing multimedia data by vectors formed from numerically captured object features. They are used in similarity search, but also, can be used for clustering and wider multimedia analysis applications. Extracting effective feature vectors for a given data type is a challenging task. Determining good feature vector extractors usually involves experimentation and application of supervised information. However, such experimentation usually is expensive, and supervised information often is data dependent. We address the feature selection problem by a novel approach based on analysis of certain feature space images. We develop two image-based analysis techniques for the automatic discrimination power analysis of feature spaces. We evaluate the techniques on a comprehensive feature selection benchmark, demonstrating the effectiveness of our analysis and its potential toward automatically addressing the feature selection problem.