Abstract
Problem. Humans can segregate binocularly viewed textured surfaces with different disparities, forming a percept of binocular transparency. Most models of stereoscopic perception cannot successfully explain the robust computation of stereoscopic depth for transparent as well as opaque surfaces. Building on a recent model of motion transparency perception (Raudies & Neumann, J.Physiol.Paris 2009, http://dx.doi.org/10.1016/j.jphysparis.2009.11.010) we propose a model of binocular stereo processing that can segregate transparent surfaces at different depths and also handle surfaces that are slanted with respect to the observer's line of sight.
Methods. Spatial correlations among model V1 orientation tuned cells calculate initial disparity estimates, which are passed to model area V2, which integrates V1 activations. Center-surround competition of disparity signals and divisive normalization of activities leads to attraction of similar disparities, a repulsion of nearby disparities and co-existence of distant disparities along a line-of-sight. Modulatory feedback helps to integrate consistent disparity estimates while dissolving local ambiguities and multiple matches at the same location.
Results and Conclusion. The model has been probed by synthetically rendered scenes with transparent slanted planes separated in depth. Such planes can be segregated in depth once disparity differences exceed a small threshold. The model can also successfully integrate smooth depth gradients. The model has been tested on stereo pairs from the Middlebury dataset (http://vision.middlebury.edu/stereo) and robustly processes opaque surfaces. When one eye's view of a region is occluded, disparity estimates are filled in from nearby positions of that region that are visible to both eyes by a feedback loop between areas V1 and V2. The model proposes how disparity sensitive V2 cells, their lateral interactions, and feedback can segregate opaque and transparent surfaces that are slanted in depth. Center-surround interactions in the disparity domain lead to fusion for binocularly matching regions and repulsion of disparity layers that are closely spaced in depth.
Supported by the Graduate School of Mathematical Analysis of Evolution, Information, and Complexity at the University of Ulm and BMBF Brain Plasticity and Perceptual Learning 01GW0763. EM was supported in part by CELEST, an NSF Science of Learning Center (NSF SBE-0354378), HP (DARPA prime HR001109-03-0001), and HRL Labs LLC (DARPA prime HR001-09-C-0011).