A modified variational method for large displacement optical flow

Abstract

We present a new way to combine the propagated flow in image pyramid and dense correspondences from descriptor matching for large displacement optical flow estimation. Because the matches and the flow propagated from the coarser level in image pyramid are possibly wrong, our method uses color-based weighted linear interpolation to reduce the wrong initial flow and alleviate over-smoothing, instead of inferring and choosing the possibly right initial value of flow. Considering the frequent violation of gradient constancy assumption and inspired by the statistic on semi-synthetic image sequences, the modified gradient term is introduced. Compared to related algorithms, the proposed approach shows competitive performance for optical flow estimation.