Perceptual Reparameterization of Material Properties

Abstract

The recent increase in both the range and the subtlety of computer graphics techniques has greatly expanded the possibilities for synthesizing images. In many cases, however, the relationship between the parameters of an algorithm and the resulting perceptual effect is not straightforward. Since the ability to produce specific, intended effects is a natural pre-requisite for many scientific and artistic endeavors, this is a strong drawback. Here, we demonstrate a generalized method for determining both the qualitative and quantitative mapping between parameters and perception. Multidimensional Scaling extracts the metric structure of perceived similarity between the objects, as well as the transformation between similarity space and parameter space. Factor analysis of semantic differentials is used to determine the aesthetic structure of the stimulus set. Jointly, the results provide a description of how specific parameter changes can produce specific semantic changes. The method is demonstrated using two datasets. The first dataset consisted of glossy objects, which turned out to have a 2D similarity space and five primary semantic factors. The second dataset, transparent objects, can be described with a non-linear, 1D similarity map and six semantic factors. In both cases, roughly half of the factors represented aesthetic aspects of the stimuli, and half the low-level material properties. Perceptual reparameterization of computer graphics algorithms (such as those dealing with the representation of surface properties) offers the potential to improve their accessibility. This will not only allow easier generation of specific effects, but also enable more intuitive exploration of different image properties.