Adaptive spatial re-weighting in stereoscopic depth perception revealed by disparity reverse correlation

Abstract

The spatial structure of binocular disparity depends on the natural scene that an observer is viewing. For example, disparities are spatially uniform (at least locally) when observers are facing a wall, but disparities are spatially variable when observers are viewing the foliage of a bush. Stereoscopic depth estimates benefit from adaptive spatial pooling: large pooling regions should be used with uniform disparity patterns and small pooling regions should be used with more variable disparity patterns. To test if the humans use adaptive pooling, we conducted a reverse correlation experiment in the disparity domain. The task was to report if the central region of a random-dot stimulus appeared nearer or farther than the zero-disparity surround. The center region consisted of nine horizontal bars (60×6 arcmin) vertically flanking the fovea. Each bar's position in depth was determined by a signal disparity (+0.34 or –0.34 arcmin depending on the trial) plus independent Gaussian disparity noise. In the first condition, disparity noise variance was spatially uniform (1.71 arcmin²). In the second condition, noise variance increased with eccentricity mimicking how disparity variance increases with eccentricity in natural viewing. The total noise variance was the same between the two conditions. In the first condition, the ideal observer uniformly weights disparities at all eccentricities. In the second condition, the ideal observer gives larger weights to more foveal disparities (reliability-based weighting). We ran ~5,000 trials per condition to estimate the weights. Results demonstrate that the ideal observer predicts the observed pattern of adaptive spatial reweighting. When disparity noise increases with eccentricity, the spatial weight sharply decreased with eccentricity. When the noise variance was uniform across space, the weights were more nearly constant. We conclude that the spatial window underlying stereoscopic depth perception is adaptively regulated based on the spatial noise structure.

Meeting abstract presented at VSS 2018