Motion Segmentation Using Convergence Properties

Motion segmentation is traditionally coupled with motion detection, where each image region corresponds to a particular motion model which accounts for the temporal changes in the region. Using the motion model to estimate the second frame from the rst frame, for example, should give a very low prediction error in the corresponding region. To relax the need for accurate motion models, it is proposed to examine the convergence of the prediction error, rather than the prediction error itself. In an iterative process of motion computation followed by computing the prediction error, those points for which the prediction error is being reduced are considered as a coherent region. This segmentation approach works well even with approximate motion models that don't eliminate the prediction error.