Controlling model complexity in flow estimation

This paper describes a novel application of statistical learning theory (SLT) to control model complexity in flow estimation. SLT provides analytical generalization bounds suitable for practical model selection from small and noisy data sets of image measurements (normal flow). The method addresses the aperture problem by using the penalized risk (ridge regression). We demonstrate an application of this method on both synthetic and real image sequences and use it for motion interpolation and extrapolation. Our experimental results show that our approach compares favorably against alternative model selection methods such as the Akaike's final prediction error, Schwartz's criterion, generalized cross-validation, and Shibata's model selector.