Unsupervised motion segmentation using metric embedding of features

Abstract

Motion segmentation is a well studied problem in computer vision. Most approaches assume a priori knowledge of the number of moving objects in the scene. In the absence of such information, motion segmentation is generally achieved through brute force search, e.g., searching over all possible priors or iterating over a search for the most prominent motion. In this paper, we propose an efficient method that achieves motion segmentation over a sequence of frames while estimating the number of moving segments; no prior assumption is made about the structure of scene. We utilize metric embedding to map a complex graph of image features and their relations into hierarchically well-separated tree, yielding a simplified topology over which the motions are segmented. Moreover, the method provides a hierarchical decomposition of motion for objects with moving parts.