Model selection is a central task in computer vision. The minimum description length (MDL) method links model selection to data compression: the best model is the one which yields the largest compression of the data. The general theoretical framework for compression is Kolmogorov complexity. MDL differs from Bayesian model selection (BMS) in that it is biased against complex probability density functions. MDL is applied to a model selection problem in computer vision. The following models are considered: background, collineation, affine fundamental and fundamental models. The experiments show that the collineation model is a good choice even for sets of image correspondences for which the 'true' model is a fundamental matrix.