Motion Estimation and Segmentation Using a Recurrent Mixture of Experts Architecture

Estimating motion in scenes containing multiple motions remains a difficult problem for computer vision. Here we describe a novel recurrent network architecture which solves this problem by simultaneously estimating motion and segmenting the scene. The network is comprised of locally connected units which carry out simple calculations in parallel. We present simulation results illustrating the successful motion estimation and rapid convergence of the network on real image sequences. 1 Introduction Motion estimation is an ill-posed problem. In other words, local motion measurements are inherently ambiguous. When the scene contains only one smoothly varying motion the ill posedness can be overcome by imposing a smoothness constraint on the solution (e.g. [Poggio et al., 1985]). The smoothness assumption, however, is not valid when the scene contains multiple motions, and imposing it leads to erroneous motion estimates especially at occlusion boundaries (e.g. [Horn, 1986]). One way to m..