Empirical Study of Intermediate Filters in Order to Improve the Optical Flow Estimation

Most used models for estimating the motion of pixels between two consecutive images are the variational models, which propose an energy to estimate the motion estimation error. The argument that minimizes that energy is the optical flow (OF) estimation of the sequence. These OF estimation models fail when large displacements and illumination changes occur. To tackle large displacements, OF methods compute the estimation in an image pyramidal scheme. In each scale, an intermediate filter is applied to eliminate outliers or noise. In this work, we considered a robust OF model and we increased the robustness of this model to illumination changes and we studied effects in OF estimation of four different intermediate filters: Gaussian(GF), bilateral (BF), bilateral extended to the gradient domain (BFG), and median filter (MF). Our experiments were performed in a subset of the state of the art MPI-Sintel dataset. Using GF, we obtained an endpoint error EPE = 5.63; for BF we obtained EPE = 7.71; for BFG we obtained EPE = 7.67 and, for MF we obtained EPE = 4.77. The best performance was obtained for the MF but the GF performs also well. Considering processing time, the GF could be also suitable for optical OF.