Abstract
Binocular disparity is a primary depth cue with an extensive psychophysical literature. However, binocular disparity has rarely been studied in context of natural viewing. Natural stereo-images are complex and highly variable relative to the artificial stimuli traditionally used to probe binocular systems. Natural stimuli associated with a given target disparity are associated with a wide variety of contrast patterns. Our previous work has shown that this natural stimulus variability has a major impact on performance. However, the impact of natural variation in local depth structure—disparity contrast—on performance has yet to be examined. Here, we investigate the joint effects of disparity magnitude and disparity contrast on disparity discrimination performance in natural stimuli. First, we sampled thousands of natural stereo-image patches from a natural image database with co-registered range information at each pixel. Image patches were binned both by disparity contrast, and by the disparity of the target (i.e. center) pixel. Disparity contrast varied from 0 to 8arcmin. Target disparity magnitude varied from 0 to 15arcmin of uncrossed disparity. In a fully-crossed 2IFC design with the sampled natural image patches, we measured disparity discrimination thresholds in three human observers. The task was to select the interval with the stimulus that appeared to be farther away. Each interval of every trial contained a unique stereo-image patch; no stimulus was presented twice. In each of five disparity-contrast conditions, we measured five five-level psychometric functions (100trl/lvl). Each psychometric function was fit with a cumulative Gaussian, from which discrimination thresholds were computed (d-prime = 1.0). The threshold data was fit with a separable two-dimensional function that was linear in disparity contrast and exponential in target disparity. Our results provide a rich picture of disparity discrimination in natural scenes.