Abstract
Deformable objects such as fabrics, rubber, and food can be distinguished by their textures, and also by the different motions they exhibit when interacting with external forces. These motion patterns can be used to estimate the intrinsic mechanical properties of the objects (Bouman et al 2013). Little is known however about human perception of mechanical properties (e.g. stiffness) in dynamic scenes. We use a dissimilarity scaling method to study how perception of mechanical properties of fabrics is related to the physical parameters of mass and bending stiffness using videos of both simulated and real fabrics. The stimuli are videos containing a hanging fabric moving under oscillating wind. In each trial, observers were shown a pair of these videos and asked to indicate on a scale of 0-100 how different the material properties of the two fabrics are from each other. In Experiment 1, we used Blender physics engine to simulate the cloth behavior. Four values are sampled for each of the 3 parameters: mass, structural stiffness, and bending stiffness. All the fabrics have the same textures. For each rendered video, the wind direction was randomized along a lateral plane. Five observers finished 2016 paired-comparisons, which were analyzed with a non-metric multidimensional scaling method to learn lower-dimensional embedding of the data. In Experiment 2, we performed the same experiment with videos of the real fabrics with similar scene settings. In the 2D perceptual embedding, we found that one dimension was best correlated with the mass, while the other dimension was correlated with the bending stiffness. Structural stiffness and wind direction did not predict perceptual dissimilarities. 2D embedding of the real fabrics also showed similar results. Together, the experiments suggest humans can estimate intrinsic mechanical properties of fabrics from dynamic scenes, while discounting variations in textures and direction of the wind force.
Meeting abstract presented at VSS 2015