Global motion segmentation for mid-level representation of moving images

In the layered image representation, expected to open up a new generation of image coding, moving images are decomposed into multiple image layers, viz. Global image regions undergoing different coherent motion which is well represented with a parametric transformation model, and then each global image region is tracked in the entire image sequence. One of the most serious difficulties with the presentation is the problem how to decompose image sequences into multiple image layers, which is referred to as the global motion segmentation problem. To solve this problem, we can take up two different approaches: an approach of gradual migration from a local optical flow representation of motion to a global segmentation representation of motion, and an approach of direct global motion segmentation which operates on image sequences straightforwardly to perform both global segmentation and estimation of parameters of multiple parametric transformation models simultaneously on the assumption that moving images are composed of multiple image layers. This paper concentrates on the direct, approach, and presents an alternately iterative grid search algorithm for it. The experiments using natural image sequences demonstrate that the direct approach works as well as we expect.