In DC Kersten and W Richards (eds.), Perception as Bayesian Inference (pp. 495-498). Cambridge: Cambridge University Press. 1996.
[Commentary on Adelson EH, and Pentland AP (1996) "The Perception of Shading and Reflectance". In Perception as Bayesian Inference, DC Knill and W Richards, eds.]
The authors admit that their approach is based on a relatively simple domain, and will have to be developed further if it is to be applied to more realistic situations. But can it really provide the basis for a better understanding of more general-purpose vision? In what follows, it will be shown that this approach is based on several rather strong hypotheses, some of which will have to be seriously modified or even replaced if an extension is to be made to more general domains. Three sets of issues are of particular concern here: the kinds of scenes that can be accurately interpreted this way, the relevance of such models to human vision, and the tradeoffs between interpretative power and processing speed.