Perception of structure from motion: is projective correspondence of moving elements a necessary condition?

Abstract

A fundamental assumption of almost all existing computational analyses of the perception of structure from motion is that moving elements on the retina projectively correspond to identifiable moving points in three-dimensional space. The present investigation was designed to determine the psychological validity of this assumption in several different contexts. The results demonstrate that the ability of human observers to perceive structure from motion is much more general than would be reasonable to expect on the basis of existing theory. Observers can experience a compelling kinetic depth effect even when the pattern of optical motion is contaminated by large amounts of visual noise (e.g., where the signal to noise ratio is less than 0.15). Moreover, the optical deformations of shading, texture, or self-occluding contours, which would be treated as noise by existing computational models, are analyzed by human observers as perceptually salient sources of information about an object's three-dimensional form. These results suggest that the modular analyses of visual information that currently dominate the literature will have to be modified if they are to account for the high level of generality exhibited by human observers.