Segmentation of Global Motion using Temporal Probabilistic Classification

The segmentation of pixels belonging to different moving elements within a cinematographic image sequence underpins a range of post-production special effects. In this work, the separation of foreground elements, such as actors, from arbitrary backgrounds rather than from a blue screen is accomplished by accurately estimating the visual motion induced by a moving camera. The optical-flow field of the background is recovered using a parametric motion model (motivated by the three-dimensional pan-and-zoom motion of a camera) embedded in a spatiotemporal least-squares minimisation framework. A maximum a posteriori probability (MAP) approach is used to assign pixel membership (background, uncovered, coveredand foreground )d efined relative to the background element. The standard approach, based on class-conditional ap rioridistributions of displaced-frame differences, is augmented by information capturing the expected temporal transitions of pixel labels.