A Multigrid Method for Motion Estimation Based on Diffusion Process

摘要

In this paper, diffusion models are investigated for optical flow estimation. In order to have efficient implementation, Weickert developed a successful Additive Operator Splitting (AOS) scheme for the diffusion filtering. There exists a tradeoff between the computational complexity and the accuracy in the AOS scheme. To achieve better estimation for the general nonlinear diffusion models, we work directly on the semi-implicit scheme instead of the modified semi-implicit scheme as in the AOS scheme to avoid the approximation. Taking both accuracy and efficiency into account, we propose the Multigrid method to solve the resulting system from the semi-implicit scheme. Experiments on real-image sequences demonstrate that more accurate estimation can be obtained.