Abstract:

A visual system makes assumptions in order to assign interpretations of shape, lighting, or motion to visual data. The assumption of \"generic view\" Koenderink79,Binford81,Biederman85,Nakayama92 states that the observer is not in a special position relative to the scene . Researchers commonly use a binary decision of generic or accidental view to disqualify scene interpretations that assume special viewpoints Lowe85a,Malik87,Richards87,Pentland90,Leclerc91,Jepson92. Here we show how to use the generic view assumption to quantify the likelihood of a view, adding a new term to the probability of a given image interpretation. The resulting framework better models what the eye sees and reduces the reliance on other prior assumptions. It may lead to computer vision algorithms of greater power and accuracy or better models of human vision. We show applications to the problem of inferring shape from a shaded image.