Geodesic-based probability propagation for efficient optical flow

Optical flow estimation has bottlenecks such as large displacement and motion blur. In this Letter, the authors propose a geodesic-based probability propagation (GeoFlow) method combining the global geodesic with local spatial similarity to build a non-local superpixel graph. To achieve efficient belief propagation, a probabilistic framework for optimising the Markov random field (MRF) objective is proposed. In this way, the limitation of local propagation can be tackled in the global image level, and the probabilistic framework reduces computational complexity in the optimisation. In experiments, their method showed promising performance by improving the results on two public large displacement benchmark datasets.