Abstract
Different objects and materials—such as shoes, starfish or ice-cream—have distinctive shapes that we readily perceive and use to recognize and categorize them. But where do those shapes come from in the first place? Objects and materials end up with particular shapes due to some kind of generative process, such as manufacture, biological growth, aging and weathering, or self-organization in response to external forces. These generative processes specify the ‘rules’ and ‘parameters’ that shape things and stuff into specific forms. Here, I suggest that when we view novel shapes, we infer a (primitive) model of the underlying generative processes and that this model facilitates us in many tasks related to shape and material perception, including: (a) identifying physical properties (e.g., viscosity, elasticity, ductility); (b) predicting future states as the sample moves and interacts with other things; (c) judging similarity between different shapes and (d) predicting what other members of the same category might look like (‘plausible variants’), even when you’ve only seen one or a few exemplars. Subjects were shown single exemplars of novel objects (‘seed object’) that were created by specific generative processes. They were then asked to rate the ‘similarity’ and ‘plausibility’ of other novel objects (‘variants’). Some of the variants were created by parametric variations of the same generative process, while others were generated by a different process. Variants that were created by the same generative process as the seed were judged as more similar and more plausible variants of the seed object than variants that were equally similar in Euclidian terms but which were created by a different generative process. This suggests generative models play an important role in shape representation. A model based on identifying statistical regularities in the shape hints at how the visual system could infer generative models from single exemplars.
Meeting abstract presented at VSS 2015