Optical flow estimation via steered-L1 norm

Abstract

Global variational methods for estimating optical flow are among the best performing methods due to the subpixel accuracy and the ‘fill-in’ effect they provide. The fill-in effect allows optical flow displacements to be estimated even in low and untextured areas of the image. The estimation of such displacements are induced by the smoothness term. The L1 norm provides a robust regularisation term for the optical flow energy function with a very good performance for edge-preserving. However this norm suffers from several issues, among these is the isotropic nature of this norm which reduces the fill-in effect and eventually the accuracy of estimation in areas near motion boundaries. In this paper we propose an enhancement to the L1 norm that improves the fill-in effect for this smoothness term. In order to do this we analyse the structure tensor matrix and use its eigenvectors to steer the smoothness term into components that are ‘orthogonal to’ and ‘aligned with’ image structures. This is done in primal-dual formulation. Results show a reduced end-point error and improved accuracy compared to the conventional L1 norm.