Aggregation Methods for Optical Flow Computation
Abstract
Global variational methods for optical flow estimation suffer from an over-smoothing effect due to the use of coarse-to-fine schemes. 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 patches. First, semi-local variational estimations are performed in overlapping square patches. Then, a global discrete optimization, based on an aggregation scheme and not prone to over-smoothing, selects for each pixel the optimal motion vector from the ones estimated at the preceding stage. The overall computation framework is simple and can be straightforwardly parallelized. Experiments demonstrate that this novel approach yields better results than the baseline global variational method: more accurate registration is globally achieved and motion discontinuities are sharpened.