From Image Statistics to Visual Inference

Dan Kersten

A major challenge for vision is the development of theories that make quantitative predictions of visual performance from natural image input. Such theories require the development of generative models of image formation tailored to the functional tasks of vision. Generative models explicitly represent the factors or causes important to estimate for a specific task. These models also make explicit the information in the image that may confound those estimates. Generative models can either be developed close to the level of the image in terms of sufficient statistcs that support inference (image-based models), or at the level of the objects, materials, and illumination that generate the image (scene-based models). I will describe recent progress in understanding human curve perception that makes use of image-based generative models. I will also show how Bayesian decision theory can be applied to scene-based models to understand how the visual system resolves ambiguity in the perception of surface material color.