Partitioning the effects of distinct natural-scene properties on visual performance

Stimulus variability is the primary source of perceptual uncertainty in everyday natural tasks. How do different properties of natural scenes contribute to this uncertainty? Using binocular disparity as a model system, we report a systematic investigation of how various forms of natural stimulus variability impact performance in a stereo-depth discrimination task. We show, using new analytical methods, the specific impact that two distinct sources of stimulus variability—luminance-pattern variation and local-depth variation—have on discrimination performance. First, we obtained a stimulus set, sampled from a natural stereo-image database with co-registered laser-based range-data at each pixel; this data set contains both natural luminance and natural depth variability. Next, we generated a second data set from the first, by flattening the local-depth profile of each stimulus; this data set contains near-identical luminance variability, but no depth variability. Stimuli were presented using a 2IFC paradigm. Each interval of each trial contained a unique stereo-image patch. The task was to report which patch appeared to be farther away. Each of three human observers collected 20,000 trials across two double-pass stereo-depth-discrimination experiments—one with each stimulus set. The stimuli in each trial were matched across passes, within and between experiments. The resulting stereo-depth discrimination thresholds indicate the absolute limits of discrimination performance at five disparity pedestals and three disparity-contrast levels. Analysis of each double-pass experiment indicates the relative importance of stimulus-variability and internal-noise in determining performance limits. And a novel quasi-quadruple-pass analysis, comparing the data from the two double-pass experiments, allows us to partition how luminance-pattern variability and local-depth variability limit performance in natural scenes. Our results show that luminance-pattern variation and local-depth variation have distinct and, surprisingly, almost completely dissociable effects on performance, and provide a rich picture of the factors contributing to human disparity discrimination performance in natural scenes.