Fast segmentation of multiple motions

Abstract

A guided sampling method for robust segmentation of multiple motions is introduced. It is substantially faster than random sampling as it effectively makes use of the spatial proximity of the points belonging to each motion. A fast high-breakdown robust estimator called Guided-LKS (GLKS) is devised using the guided search to minimize the k-th order statistics of squared distances. A number of experiments on homography estimation problems are presented. They involve up to eight motions and benchmark the performance of GLKS estimator in comparison to a number of state of the art robust estimators. The results show that while GLKS performs similar to other estimators in terms of segmentation accuracy, it significantly outperforms them in terms of computation time. The fast convergence and high breakdown point of GLKS make this estimator an outstanding choice for real-time estimation and segmentation of multiple motions.