Abstract
We routinely interact with a wide range of materials. Through visual inspection, we can classify them into categories (e.g., rock), as well as infer their diverse optical (e.g., translucency) and mechanical properties (e.g., rigidity). To investigate if and how human material categorization affects the estimation of material properties, we developed a framework to systematically create images of ambiguous materials. Specifically, by training an unsupervised image generation model (StyleGAN) with transfer learning, we obtained models that synthesize images from three material classes: soaps, rocks, and squishy toys. Via linear interpolation of models’ latent spaces and weights, we can smoothly morph one material to another. We sampled ten morphing sequences in which a soap is gradually transformed into a rock and then into a squishy toy in 13 steps. In Experiment 1, ten participants rated each image on five attributes: translucency, glossiness, surface smoothness, rigidity, and brittleness. In Experiment 2, the same participants performed a 10-AFC task of material identification on the same set of images. We found that estimations of mechanical and tactile properties (e.g., rigidity, brittleness, and smoothness) were modulated by morphing. Notably, the rigidity ratings gradually increased along the morphing from soap to rock, followed by a decrease from rock to squishy toy. In contrast, optical properties (e.g., translucency, glossiness) were not correlated with morphing. Finally, participants were uncertain about the material identity of images close to the midpoint of cross-material morphing, sometimes perceiving them as entirely different materials like jelly, candy, wax, and glass. Such morphed materials with high category ambiguity also show high variance in estimations of mechanical properties. Together, our results suggest that material categorization significantly impacts the inference of mechanical properties, especially when material identity is ambiguous.