Abstract
A fundamental issue in understanding visual object recognition concerns the structure of 3-D shape representations. One influential, and longstanding. view is that these representations consist of structural descriptions based on volumetric part primitives such as geons (e.g., Biederman, 1987). This study reports new converging evidence from experimental psychology, and neural network modelling, that undermines this hypothesis. Empirical data from studies using repetition priming, and whole-part matching, paradigms show that volumetric part boundaries have no special status in shape representations. Instead, we outline a surface-based model in which 3-D object shape representations consist of spatial configurations of 2-D surface patches. Some aspects of these representations are modelled using associative networks that learn patterns of inter-correlations among spatially adjacent surface patches by simple Hebbian learning. The resulting representations show several surprising properties including an emergent volumetric part structure. This model provides an account of the current empirical data, as well as showing that volumetric structure can be encoded in surface-based representations that do not contain any volumetric primitives such as geons. It is argued that this surface-based hypothesis provide a new framework for understanding shape representation in human vision.
This research was funded by ESRC (UK) project grant R000239512.