The resolution of perceptual ambiguity in complex scenes

The interpretation of images from both simple and complex scenes is ambiguous because of rendering and geometric projection. By studying human perception at the edge of ambiguity, we can gain insights into the constraints used by the visual system, and begin to bridge the gap between the simple images of the laboratory and complex natural scenes. Computer graphics provides the means to generate photo-iealistic images that produce strong perceptual interpretations, yet are theoretically indeterminate. I will show several illusions that illustrate the decisions human perception makes about object attributes and relations. In particular, I will show how the addition of complexity in the form of multiple light sources, with corresponding multiple shadows, reduces perceptual ambiguity regarding movement in depth of the casting object. How does the visual system make decisions about the properties of objects and their relations? The theory of Bayesian inference provides a natural framework to understand how to resolve ambiguity. Bayes provides a theory of rational inferences about hypotheses, given data, in the face of uncertainty. But there is more to Bayes than a priori modeling of hypotheses. A recurring theme over the years has been to view perception as a process of model construction (e.g. Barlow, Mumford). An appropriate model to "explain" the image data must be selected based on plausibility and utility (Yuille). Recent results by Freeman, MacKay and others have shown that Bayesian analysis provides the basis for model selection in terms of robustness and economy of the model description. This analysis can be applied to understanding the resolution of depth ambiguity given cast shadows, and promises to provide fruitful insights into the perceptual processing of both simple and complex scenes at early and later levels of computation.