Abstract
Since the inventions of the random-dot stereogram, empirical and theoretical studies of stereopsis have focused on understanding how correspondence is achieved and binocular disparity is computed. Correspondence computations are thought to give rise to a disparity “map” that makes explicit the disparities of all points common to the two retinal images. It has been widely assumed that there is a simple and direct relationship between the disparity map and the perception of stereoscopic depth. In this view, all that is needed to transform the disparity map into a representation of egocentric distance is the choice of an appropriate scaling transformation. Here, a theoretical analysis is presented that demonstrates that this view is incorrect. It is argued that two principles are needed to understand how the visual system computes surface properties from the pattern of binocularly matched contrast signals. First, it is shown that occlusion geometry imposes a constraint on the polarity of image contrasts that can form in the two eyes, and that this constraint restricts the ordinal depth and lightness relationships that can be attributed to contrast signals. Second, it is shown that the relative magnitudes of image contrast are used to infer the opacity (or transmittance) of occluding and transparent surfaces. Thus, the polarity and magnitude of binocularly matched image contrast play distinct roles in stereoscopic surface perception, and it is argued that these distinct roles can be captured by two theoretical principles of perceptual inference. These two principles provide a unified framework for understanding how the visual system infers the relative depth, lightness and opacity of stereoscopic surfaces from the contrast relationships that occur along depth discontinuities.