Abstract
Recovering 3D shape from shading alone is an ill-posed problem, so the observer must make assumptions about the lighting, surface material, and viewpoint. Typically, the assumed light position is above and slightly to the left (Sun & Perona, 1998). Consider a plane containing the eye, the point of interest on the object, and the assumed light position. The plane would be tilted slightly counter-clockwise from vertical to contain the above-and-left preference. We asked where in this plane is the assumed position of the light. We did so by changing the angle of incident light on virtual 3D objects and determining the accuracy of observer percepts for each angle. The objects were smooth, abstract 3D shapes with local Lambertian shading. They were viewed monocularly. The angle in the above-mentioned plane of the distant point light source was varied across trials over a range of +/−66 deg (above/below) from the viewpoint. On each trial, the observer indicated local surface orientation at a marked point on the object by rotating a gauge figure to appear normal to the surface (Koenderink et al., 1992). Observer settings were more accurate when the light position was above rather than below the viewpoint. The angular difference between the indicated and actual surface normal was minimized when the light was 20–30 deg above the viewpoint. Measurements of slant and tilt errors, as well as the variance of the settings, supported this result. These findings confirm that the human visual system has an above-and-left prior for light position and show that the preferred position is 20–30 deg above the viewpoint.
NIH, NSF, Microsoft Research, The Okawa Foundation, and The Alfred P. Sloan Foundation.