Robust Flow Estimation With Global Sample Consensus

Most of leading optical flow methods involve a step to estimate dense flow from putative sparse matches.Although RANSAC(RANdom SAmple Consensus)is known to be robust against significant outliers,so far it can be used for only parametric model estimation,and is prohibited for flow estimation due to the high dimensionality of the flow space.In this paper,we propose a robust flow estimation method based on the hypothesize-verify framework of RANSAC.We first formulate flow estimation and outlier filtering as an optimization problem maximizing a flow score function.The solution of which,however,is NP-hard,and meantime involves a sparse-to-dense flow interpolations process.We then introduce an efficient estimation method that reduces the complexity to be sub-linear.Input matches are first grouped and ranked,then a progressive procedure is invoked to separate inliers and outliers,which requires only several times of flow interpolations.The flow score can be computed with any robust matching measure,and the advantages over previous energy-based methods and piecewise methods are demonstrated.We call the proposed method as FlowSAC,and reveal its connections with previous RANSAC-family methods.Experiments on Sintel and Kitti datasets showed that FlowSAC constantly outperform previous outlier filtering methods.It produces more inliers and less outliers,and presents to be apparently less sensitive to the increase of outlier ratio.