This paper presents an extensive survey of model selection techniques for computer vision applications. A large number of existing model selection criteria and a new model selection criterion (SSC) are introduced and their performance for two important computer vision tasks: motion estimation and range segmentation are evaluated and compared. Various factors affecting the performance of different criteria are introduced and their effects are compared by virtue of conducting controlled experiments using synthetic and real data. Our results show that the performance of different model selection criteria are affected by the size of data and the amount (and distribution) of noise as well as of complexity of models used in an application.