Performance issues in shape classification

Abstract

Shape comparisons of two groups of objects often have two goals: to create a classifier to separate the groups and to provide information that shows differences between classes. We examine issues that are important for shape analysis in a study comparing schizophrenic patients to normal subjects. For this study, non-linear classifiers provide large accuracy gains over linear ones. Using volume information directly in the classifier provides gains over a classifier that normalizes the data for volume. We compare two different representations of shape: displacement fields and distance maps. We show that the classifier based on displacement fields outperforms the one based on distance maps. We also show that displacement fields provide more information in visualizing shape differences than distance maps.