Stability and instability of perceived shape from shading

Abstract

Images of an object can vary considerably due to variation in illumination. Despite this, humans are surprisingly adept at recovering stable estimations of shape, suggesting the presence of invariant structure in images upon which our perception relies. Failures of shape constancy also occur where a change in one factor (e.g., illumination) is misperceived as a change in another (e.g., shape). This thesis examines conditions predictive of success and failure of shape constancy and identifies image structure that corresponds with our perception. Our experiments suggest that specific alignment between surface geometry and illumination results in misperception of local and adjacent shape features. Anisotropic shape features, for example a cylindrical ridge, produces minimal shading when the illuminant is aligned with its axis of minimal curvature, causing its apparent deletion. This local deletion of shape was found to produce illusory segmentation of adjacent surface geometry. We also present examples of non-affine changes to perceived shape beyond an affine reduction of perceived relief of anisotropic shape features related to increased contrast of shading gradients produced by slant variation parallel to the illumination direction, and reduced contrast produced by slant variation in the perpendicular direction. The addition of specular reflections reduced the size of the effects, however the pattern of shape misperception was preserved. These effects were also observed when rendered using a light map containing multiple, distributed strong light sources, suggesting that the perception of Lambertian surfaces is primarily determined by low frequency structure of the illuminant. We conclude that illumination affects shape perception in a complex, piecewise manner and propose directions for future research.