Abstract
Detection of spatial targets is a fundamental visual task. As with laboratory synthetic stimuli, performance in natural backgrounds depends on multiple dimensions. Using a constrained sampling approach (sorting millions of gray-scale natural background patches into narrow bins along multiple dimensions) Sebastian et al. [1] found that the thresholds of template-matching observers are the separable product of the local luminance (L), contrast (C), and phase-invariant similarity (S) of the natural background (similarity was the cosine similarity between the amplitude spectrum of the background and target). They also showed that for each dimension alone, human thresholds are consistent with this prediction from natural scene statistics. Here we tested whether human thresholds are consistent with the prediction of separability for the dimensions of contrast and similarity. Specifically, we parametrically measured threshold for an additive Gabor target (4 cpd) for 3 levels of contrast X 5 levels of similarity. These levels of contrast and similarity spanned the ranges observed in the natural backgrounds, with the restriction that the thresholds could be measured without significant clipping in the display. Thresholds for each bin were estimated from psychometric functions, where the natural backgrounds were randomly sampled from the bin without replacement. The level of luminance in the present experiment was approximately twice that in the Sebastian study, and three subjects from the Sebastian study ran in the current study. Combining the data across the two studies, we found (in agreement with the natural scene statistics) that human thresholds were approximately consistent with separable Weber's law for luminance, contrast and similarity: threshold = k0 x (L+k1) x (C+k2) x (S+k3). Thus, human detection performance appears to be predicted (up to a single scale factor) directly from the natural-image statistics over much of the space of natural background patches. [1] Sebastian, Abrams and Geisler, PNAS, 2017
Meeting abstract presented at VSS 2018