Abstract
Anti-correlated random dot stereograms (AC-RDS) drive disparity-selective cortical neurons, elicit vergence eye movements, but fail to give rise to perception of depth (Cumming and Parker, 1997; Masson et al., 1997). Stimulation with AC-RDS elicits V1 disparity tuning curves and vergence movements that are inverted in sign, consistent with quasi-linear binocular combination. Recently, we have described a class of anti-correlated RDS that does give rise to a compelling sense of reversed depth. This stereogram is composed of local triads (two dots matching one, similar to the classical Panum's limiting case (PLC)), with one element having reversed contrast sign. We describe a model that accounts for the differences in perception observed with standard and PLC anti-correlated stereograms. The model uses standard binocular energy cells at several spatial scales in stage one, followed by within-scale inhibition between neurons differing by one half wavelength (pi) in disparity-tuning, followed by a disparity-specific weighted summation across scales. The within-scale inhibition greatly diminishes non-specific noise including the response to monocular inputs, and the multi-scale integration emphasizes features with multi-scale disparity consistency. In the case of standard AC-RDS, the stage one disparity estimates diverge with scale and the multi-scale model produces weaker responses. In contrast, in the case of the PLC AC-RDS, the stage one estimates show less divergence, and respond over a smaller range of scales. The output of the second stage, consistent with perception, is reversed. The model also responds suitably to transparent surface RDS and crisply represents depth discontinuities. Finally, the model predicts that a population of disparity selective neurons beyond V1, possibly in V2, which because of multi-scale integration, will be found that exhibits significantly broader spatial frequency tuning than is seen in disparity-selective V1 cells.