Abstract
Visual estimation of material properties is an ill-posed problem: an object's image varies with surface reflectance, but also with illumination and 3-D surface geometry. Previous studies have shown that perceived gloss depends on 3-D surface geometry: bumpier surfaces are often perceived to be glossier. In addition, surface gloss can modulate perceived shape. Are shape and reflectance jointly estimated? Here we ask whether human perception reflects a generative model such that errors in one property predict errors in the other. To this end, we determined quantitative estimates of perceived shape and perceived gloss for the same stimuli. We generated two sets of physical objects: one varied in gloss and one in bumpiness (depth extent). Mixtures containing different proportions of matte and glossy varnish were applied to spheres resulting in eleven gloss levels. To estimate the reflectance characteristics of these objects, we determined the surface reflectance parameters of the Ward model that minimized the difference between photographs of the objects and rendered images. 3D printing produced eleven arrays of random-height ellipsoids varying in bumpiness (depth range). Observers reported the perceived glossiness and bumpiness of computer-rendered ellipsoid arrays by referencing the physical objects. Glossiness and bumpiness of the rendered objects varied independently across trials. Rendered stimuli were viewed monocularly without head movements. In contrast, the physical objects, lit with ambient and direct light, were viewed binocularly with unrestricted viewing. Gloss judgments increased as a function of both physical gloss and physical bumpiness. Although perceived bumpiness was largely independent of physical gloss, bumpiness was increasingly underestimated as depth increased. Thus, observers indeed estimate shape and gloss jointly: underestimation of shape is coupled with overestimation of gloss, consistent with the effects of these two variables on images of physical objects.
Meeting abstract presented at VSS 2016