Anisotropic Huber-L1 Optical Flow

Proceedings of British Machine Vision Conference (BMVC)

The presented work is motivated by the problem of restoring severely degraded historic video material via an optical flow-based interpolation. In order to increase the robustness as well as the accuracy of discontinuity preserving variational optical flow models, we propose two extensions. First, we will discuss the deficiencies of an isotropic Total Variation regularization and introduce an anisotropic (i.e. image-driven) regularization based on the robust Huber norm. Second, to cope with the gross outliers contained in historic video material, we propose a novel spatio-temporal regularization approach. Instead of assuming gradual flow changes over time, we impose a symmetry constraint with respect to a central frame. The benefit of the suggested enhancements is illustrated qualitatively on historic video material and quantitatively on the Middlebury optical flow benchmark.

PDF