A generalisation of model selection criteria

Abstract

In this article we generalise some of the existing model selection criteria used in statistics and computer vision. Model selection criteria are mostly used to decide which model is more appropriate for explaining a specific data set. We adapt these criteria in a way that they can be used for the evaluation of models fitted to different data sets and for the evaluation of sets of models. We found this adaption especially useful in situations where we have a set of models with domains that may overlap and where we need to find an optimal subset of models for explaining the whole data set. Adapted model selection criteria were then succesfully used in range image segmentation application (reverse engineering), based on the recover-and-select paradigm.