Modern large displacement optical flow algorithms usually use aninitialization by either sparse descriptor matching techniques or denseapproximate nearest neighbor fields. While the latter have the advantage ofbeing dense, they have the major disadvantage of being very outlier prone asthey are not designed to find the optical flow, but the visually most similarcorrespondence. In this paper we present a dense correspondence field approachthat is much less outlier prone and thus much better suited for optical flowestimation than approximate nearest neighbor fields. Our approach isconceptually novel as it does not require explicit regularization, smoothing(like median filtering) or a new data term, but solely our novel purely databased search strategy that finds most inliers (even for small objects), whileit effectively avoids finding outliers. Moreover, we present novel enhancementsfor outlier filtering. We show that our approach is better suited for largedisplacement optical flow estimation than state-of-the-art descriptor matchingtechniques. We do so by initializing EpicFlow (so far the best method onMPI-Sintel) with our Flow Fields instead of their originally usedstate-of-the-art descriptor matching technique. We significantly outperform theoriginal EpicFlow on MPI-Sintel, KITTI and Middlebury.
展开▼
