Computer Vision News - August 2016

Computer Vision News Research 19 Research Conclusions Several challenges in the estimation of optical flow from real-world videos have not been solved yet: the main ones are occlusions, motion discontinuities and large displacements, all challenges which are present in real-world videos. Models addressing the issue of small displacements have been proposed with some success, but these models do not provide an effective answer to the large displacements stemming from fast-motion video. Estimation modelling of large displacements has been based to date on coarse-to-fine optimization schemes, a technique of energy minimization whose main flaw is that it may generate error- propagation and whose convergence is quite difficult to guarantee. The EpicFlow paper therefore suggests to initiate the optical flow estimation by interpolating a sparse set of matches in a dense manner. It then uses this estimate to initialize a one-level energy minimization and obtain the final optical flow estimation. Best published results! Results Results on the MPI-Sintel test set are given in the table below. Parameters are optimized on the MPI-Sintel training set. EpicFlow outperforms the state-of-the- art with a gap of 0.5 pixels in AEE compared to the second best performing method (TF+OFM) and 1 pixel compared to the third one (DeepFlow). In particular, the results are improved for both AEE on occluded areas and AEE over all pixels and for all displacement ranges. In addition, EpicFlow's approach is much faster than most other methods.