Empirical Analysis of Descriptor Spaces and Metrics for Image Classification

The objective of this work is to devise an approach for measuring similarity of low-level content features in multi-descriptor spaces. Since no single descriptor is able to represent all the properties and patterns encapsulated in natural images, the combination of several descriptors appears to be a sensible strategy to increase their discrimination power and classification properties. We are interested in combining descriptors using weights according to their discriminative properties for given semantic concepts. In an attempt to better understand the discriminatory properties of relevant descriptors a thorough empirical and statistical analysis of their basic properties is reported in this paper. The goal is to capture the behavior of descriptors by considering the particular syntax and appropriate distance metric depending on the syntax of each descriptor.