|Stochastic Systems Group|
Helmholtz, who set the foundations for the study of visual perception, said: what we perceive is our best guess as to what is in the world, based on the current sensory data plus our prior experience. He referred to this process as "unconscious inference," and it can be thought of as model-based vision utilizing scene statistics. There are two flavors, which might be called vision-as-deduction and vision-as-estimation. The second one, vision-as-estimation, is increasingly popular, and indeed "Bayesian" has become a common buzzword in both human vision and machine vision. In my lab, we have found that the vision-as-estimation stance is useful in several domains, including motion perception, lightness perception, and material perception. I'll describe some of the progress we have made, and some of the problems that remain.
Problems with this site should be emailed to firstname.lastname@example.org