A comparative study of model selection criteria for computer vision applications

During last three decades many model selection techniques have been developed, many of those have also been employed in computer vision applications. Interestingly, most of those criteria are based upon assumptions that are rarely realised in practical applications. As a result, the question of which model selection criterion works best for a particular application is of interest to many computer vision researcher and practitioners alike. This paper is an attempt to provide a satisfactory answer to this question for some well-known computer vision applications. Here, we present a comparative study of a large number of the existing model selection criteria for three computer vision tasks including: range modelling, motion modelling and merging of 3D surfaces in range data. Compared with other criteria, the results show that the surface selection criterion (SSC) appears to perform generally better for the above applications.