Semi-local variational optical flow estimation

Global variational methods for optical flow estimation usually suffer from an over-smoothing effect. We propose a semi-local estimation framework designed to integrate and improve any variational method. The idea is to implicitly segment the minimization domain into coherently moving windows. In a first time, local variational estimations are performed in overlapping candidate square regions. Then, a global discrete optimization, non subject to the over-smoothing introduced by variational approaches, selects the optimal window for each pixel. Experimental results show an increasing of the sharpness of discontinuities and a significant improvement of global registration errors compared to the results of the baseline global variational method.