We are evaluating various optical flow algorithms. By now our favourite algorithm is the DIS optical flow in OpenCV, if computational efficiency is not taken into consideration. NVOF is significantly faster than DIS, but it seems the accuracy of DIS is better. We observe that NVOF generates flow fields with jagged motion boundaries and noisy blocks even for the slow preset, while the DIS flow is quite smooth. DIS has a published paper, and the errors reported in the paper is also lower than NVOF as reported in this link. As described in the paper of DIS, a notable point is the variational refinement step significantly reduces the PEP error (20% lower). Maybe NVOF does not use such a variational refinement step? Will the variational refinement be implemented in the future?