A Bayesian approach to binocular stereopsis

摘要: We develop a computational model for binocular stereopsis, attempting to explain the process by which information detailing 3-D geometry of object surfaces is encoded in pair stereo images. design our within Bayesian framework, making explicit all assumptions about nature image coding and structure world. start deriving formation, introducing definition half-occluded regions simple equations relating these disparity function. show that function alone contains enough determine regions. use relations derive formation are explicitly represented computed. Next, we present prior series three stages, or “worlds,” where each world considers an additional complication prior. eventually argue must be constructed from local quantities scene geometry-i.e., depth, surface orientation, boundaries, creases. In addition, new dynamic programming strategy estimating quantities. Throughout article, provide motivation development psychophysical examinations human visual system.