We propose a computational model for the perception of surfaces from binocular disparity within a Bayesian framework. The model attempts to reconstruct the perceived surfaces from their differential structure, which is concurrently estimated from initial data of given binocular disparity. To achieve this, the model consists of four processes: a continuous-valued process that represents the surface depth, and three binary processes that represent the presence or absence of depth discontinuities, nonzero second spatial derivatives of the surfaces, and outliners in given disparity data. By strongly coupling these four processes, the perceived surface is reconstructed within regions bounded by salient stereo features such as the discontinuities and nonzero second spatial derivatives of the surface, while rejecting outlying points from the surface. We show through simulations that the model is consistent with some psychophysical experiments that address the perception of stereoscopic surface.
展开▼
