Combinatorial Regularization of Descriptor Matching for Optical Flow Estimation

Abstract

One fundamental step in many state of the art optical flow methods is the initial estimation of reliable correspondences. It is well-established to extract and match features such as HOG to handle large displacements. We propose a combinatorial refinement of the initial matching. Optimization is done in the space of affine motion, where we regularize between neighboring points and similar regions. The evaluation on the MPI-Sintel dataset shows that the proposed method removes outliers from the initial matching and increases the number of reliable matches. The proposed refinement improves all optical flow algorithms that build upon pre-computed correspondences.