Segmentation-Based Motion with Occlusions Using Graph-Cut Optimization

Abstract

We propose to tackle the optical flow problem by a combination of two recent advances in the computation of dense correspondences, namely the incorporation of image segmentation and robust global optimization via graph-cuts. In the first step, each segment (extracted by colour segmentation) is assigned to an affine motion model from a set of sparse correspondences. Using a layered model, we then identify those motion models that represent the dominant image motion. This layer extraction task is accomplished by optimizing a simple energy function that operates in the domain of segments via graph-cuts. We then estimate the spatial extent that is covered by each layer and identify occlusions. Since treatment of occlusions is hardly possible when using entire segments as matching primitives, we propose to use the pixel level in addition. We therefore define an energy function that measures the quality of an assignment of segments and pixels to layers. This energy function is then extended to work on multiple input frames and minimized via graph-cuts. In the experimental results, we show that our method produces good-quality results, especially in regions of low texture and close to motion boundaries, which are challenging tasks in optical flow computation.