Abstract
Visually estimating mechanical properties, such as stiffness or elasticity, is computationally challenging. A deformable object's shape and motion depends not only on its internal properties, but also on the external forces applied. Thus, to infer the object's properties, the visual system must somehow disentangle the causal contributions of multiple factors. To investigate how the brain achieves this, we simulated and rendered 20 short animations of rigid objects interacting with a non-rigid target object. We varied the type of interaction ('scene') as well as the target's stiffness and elasticity, i.e. whether the deformation was permanent (plastic) or the object returned to its original shape (elastic). Fifteen observers rated the apparent softness, elasticity and deformation of the targets. Despite large stimulus variations across scenes, responses were broadly in accordance with the simulated internal properties, although plastic objects were perceived softer than equally stiff elastic objects (presumably because plastic deformations were perceived larger than elastic deformations). Indeed, there was a strong correspondence between perceived stiffness and perceived deformation. We characterized the physical deformation of the objects by measuring seven deformation features (e.g. wobbling, stretching, curving) on the underlying 3D-meshes. Strikingly, we found that representing the stimuli in this 7-dimensional feature space systematically organized the stimuli by their internal properties, compensating for the effects of extrinsic factors. A linear combination of the features predicts softness perception very well (r = .93). Next, we simulated >100.000 animations in which we varied the scene layout, type and amount of external force, and the stiffness and elasticity of the target objects. This dataset allowed us to select stimuli with diverging predictions from different features to test the contribution of individual predictors and validate our model with new stimuli. The results suggest the brain achieves 'softness constancy' by representing deforming objects in a multidimensional feature space.
Meeting abstract presented at VSS 2018