A Variational Optical Flow Model for Accurate Motion Estimation from Rotational Image Sequences

There are several optical flow models which aim to capture spatial characteristics of the flow such as divergence and curl. However accurate estimation is often a key challenge. In this context, we propose a variational optical flow model motivated by the harmonic-constraint based regularization for improving the estimation of rotation in motion. Our model is the standard Horn and Schunck model with an additional constraint penalizing the curl of the flow with an anisotropic weight term. Our implementation scheme is based on the robust Chambolle-Pock primal-dual algorithm. Further, to refine the flow-edges, we use the heuristic of weighted median filtering in our algorithm as a post-processing step. The results of our model indicate that it is comparable to, and in some cases, outperform some of the top-performing variational methods across multiple datasets in both angular and end-point errors.