Abstract
The human visual system did not evolve in an ecosystem of metallic stimuli, nonetheless, we are highly accurate in our assessment of metal objects from vision alone. To evaluate the predictive power of local image statistics for observer metallicity judgement, we used computer-generated natural images--natural in the sense that they were rendered using physically-based models of plausible materials and environmental illumination. Low-level image statistics were manipulated indirectly by varying the rendering parameters of metal roughness, which blurs the reflected environment image, and the irregularity of a transparent coating, which effectively introduces a local disarray vector field. In a conjoint measurement task, both metal roughness and coating bumpiness were found to determine observer estimates of metallicity, combining in either additive or multiplicative ways for individuals. Fourier and steerable pyramid analyses show that both physical parameters of the stimulus space have predictable effects on image statistics across different object viewing angles, most notably on local statistics not represented in the Fourier power spectrum. Decision-making models using the activations of oriented filters at two levels of the steerable pyramid (with just two free parameters) replicated observer data for suprathreshold scaling of metal roughness and coating bumpiness, and could also accommodate individual differences in estimations of metallicity. We conclude that low-level, local image statistics, as represented in steerable pyramid analysis, can be used to achieve robust material perception, with individual differences accounted for by differently weighted combinations.