Perception of shape in two and three dimensions

点击翻译按钮获取中文摘要

The human visual system enables us to perceive the three-dimensional (3D) shapes of objects from the two-dimensional (2D) images on our retinas.  One of the most important kinds of visual information about object shape is image shading:  the variation of intensity from place to place in an image, due to differences in the orientation of local surface patches relative to a light source.  The problem of recovering object shape from image shading is difficult, because it is underdetermined:  any given image could have been produced by any of an infinite number of arrangements of lighting, 3D surfaces, and surface reflectance patterns.  Consequently, to perceive 3D object shape from image shading, any natural or artificial visual system must rely on assumptions, at least implicitly, about what 3D interpretations of 2D images are plausible or implausible.      The proposed research will investigate what assumptions the human visual system makes about natural scenes in order to perceive 3D shape from 2D images, and how these assumptions compare to the optimal assumptions for natural scenes.      The component projects are as follows.  First, we will formulate a statistical description of natural lighting, which will answer questions such as how direct or diffuse natural lighting tends to be.  For this purpose we will construct a novel multidirectional photometer, that measures light intensity in many directions at once.  Second, we will formulate a statistical model of natural surface patches, which will describe the shapes and surface reflectance patterns of small surface patches that are typically found in natural scenes.  This model will be based on 3D digital scans of a large number of randomly selected objects.  Finally, we will run psychophysical experiments to measure human performance in 3D shape perception tasks, using visual stimuli that either match the illumination and surface statistics that we will have measured, or depart from these statistics in various ways.  From the pattern of human performance across these tasks, we will infer what properties of the world human observers rely on to accurately perceive 3D shape.