Abstract
The distribution of luminances in natural scenes varies greatly depending on lighting, material properties, and other factors, but some luminance statistics—particularly higher-order statistics—appear to be regular. Natural scene luminance distributions typically have positive skew, and for single objects, there is evidence that higher skew is a correlate (but not a guarantee) of glossiness. Skewness is also relevant to aesthetics: it is a good predictor of fruit freshness, and preference for glossy single objects (with high skew) has been shown even in infants. Given that primate vision appears to efficiently encode natural scene luminance variation, and given evidence that natural scene regularities may be a prerequisite for aesthetic perception in the spatial domain, here we ask whether humans in general prefer natural scenes with more positively skewed luminance distributions. If humans generally prefer images with the higher-order regularities typical of natural scenes and shiny objects, we would expect this to be the case. By manipulating luminance distribution skewness (holding mean and variance constant) for individual natural images, we show that in fact preference varies inversely with skewness. This finding holds for both artistic landscape images and images from natural scene databases, including scenes with and without glossy surfaces, as well as for noise images. Across these conditions, humans prefer images with skew near zero over higher positive skew images. These results suggest that humans prefer images with luminances that are distributed relatively evenly about the mean luminance, or images with similar amounts of light and dark. Following evidence from prior brain imaging studies, we propose that human preference for higher-order statistical regularities reflects an efficient processing advantage of low-skew images over high-skew images. We further propose that human artwork generally possesses low skew luminance histograms in part as a way to achieve efficient visual processing.
Meeting abstract presented at VSS 2013