Bayesian Models for Seeing Shapes and Depth

We review computational models of shape and depth perception and relate them to visual psychophysics. The Bayesian approach to vision provides a fruitful theoretical framework both for modelling individual modules, such as stereo, shading, texture and occlusion, and for integrating their information. In this formalism we represent depth by one, or more, surfaces with prior probabilities for surface shape, corresponding to natural constraints, in order to avoid the ill-posedness of vision. On theoretical grounds, the less information available to the module (and the less accurate it is) then the more important the priors become. This suggests that visual illusions, and biased perceptions, will arise for scenes for which the priors are not appropriate. We describe psychophysical experiments which are consistent with these ideas. For integration of different modules we advocate strong coupling, so that the modules can interact during computation and the priors can be modified. This framework is rich enough to accommodate straightforwardly both consonant and contradictory cue integration and different psychophysical experiments can be understood within the Bayesian approach