Abstract
Systematic errors in laboratory psychophysics tasks can reveal sophisticated perceptual mechanisms optimized for natural-image statistics. Here, we report an examination of the human ability to average spatially varying signals in two visual domains: brightness and stereoscopic depth. The spatially varying signals were presented with nine horizontal bars stacked vertically in a one-degree area centered on the fovea. Human observers discriminated the sign of the average brightness or depth. Previously, we reported that human behavior departs notably from ideal-observer behavior in these simple spatial averaging tasks. Additionally, with identical stimuli, human errors were partly systematic and repeatable. In the current study, we show that nonlinear encoding of individual feature values prior to averaging contributes to the systematic deviation between human and ideal performance. In both brightness and depth tasks, the estimated nonlinearity was compressive across observers. By simulating the fitted models, we found that models with compressive encoding yield better discrimination thresholds than equivalent models with linear encoding when: i) additive late noise corrupted averaging and ii) feature values had small spatial variability. In models of many perceptual tasks, the addition of late noise is common. In natural images, the sample variance of luminance or binocular disparity within a local patch is biased toward small values. Indeed, simulations of the fitted models show that compressive encoding mitigates the performance drop caused by the late noise. Critically, this performance mitigation was more pronounced when the spatial variability of feature values skewed toward zero, as seen in natural images. Therefore, the systematic deviation between human and ideal observers in our laboratory averaging tasks may, at least in part, reflect the visual system’s adaptation to natural-image statistics in the presence of late noise.