Computer Vision News - August 2016

16 Computer Vision News Research Research EpicFlow Every month, Computer Vision News reviews a research from our field. This month we have chosen to review EpicFlow: Edge-Preserving Interpolation of Correspondences for Optical Flow , a research paper presenting a novel approach for optical flow estimation. Targeted at large displacements with significant occlusions, the paper suggests an approach which is fast and robust for large displacements. We are indebted to the authors (Jerome Revaud, Philippe Weinzaepfel, Zaid Harchaoui and Cordelia Schmid) for generously allowing us to use their images to help illustrate this review. The full paper (presented at CVPR 2015) is here and the source code is here . Background, motivation and novelty Accurate estimation of optical flow remains a challenging problem, despite the abundant literature on the topic. The main challenges include occlusions, motion discontinuities and displacements . A typical approach to tackle these issues is to transform the problem into an energy minimization framework with coarse-to- fine optimization. However, due to the complexity of the minimization, such methods get stuck in local minima and may fail. Researchers suggest using methods for matching points between images (e.g. HOG) to guide the coarse-to- fine optimization to obtain a full-scale dense flow field guided by the matches between the two images. However, the major drawbacks of coarse-to-fine based matching points are: (1) edges and matches are not well-defined at coarse scales (thin parts might merge); (2) coarse-to-fine matching is often unable to recover from errors that appeared at coarser scales. “ EpicFlow suggests that there may be better initialization strategies than the well-established coarse-to-fine scheme ” The main ideas of EpicFlow are: 1. Directly initialize the variational minimization with a dense interpolation of the matching computed at full image scale. 2. Leverage the fact that motion boundaries and image edges coincide most of the time. The EpicFlow three main contributions are: • A novel sparse-to-dense interpolation scheme of matches based on an edge- aware distance. • An approximation scheme for the edge-aware distance, leading to a significant speed-up without loss of accuracy. • State of the art results on MPI-Sintel and on par performance on Kitti and Middlebury.