There are two key elements in defining the problem of visual perception.
The first is that useful information about the world, such as the shape,
material, illumination, and spatial relationships of objects, is encrypted
in the image. This suggests that seeing is a process of decoding image
information. Second, the encryption process, of going from a description
of the world to an image, is not in general reversible--image information is
typically ambiguous about its causes in the scene. This second fact
suggests that vision's decoding is fundamentally statistical. Yet, human
vision has solved the problem of making reliable guesses from ambiguous
image measurements. How does it do this?

I will describe results from three different lines of work that pertain to this
question. First, by using computer graphics to create images from synthetic
scenes, we can gain insights into the general constraints used by the human
visual system to decode image information. Second, I will show how the tools
of Bayesian decision theory apply naturally to understanding and modeling how
the visual system resolves ambiguity. Third, I will describe some preliminary
brain imaging results that suggest an important role for primary visual cortex
in resolving ambiguity.