Computational Approaches to Human Shape Representation

摘要

The primary goal of this project has been a computational investigation into theunderlying representations used in human object recognition. To this end, we began with a relatively new approach to object representation in computer vision, that of aspect graphs (Koenderink van Doom, 1979). An aspect graph representation is a complete representation of an object at all image resolutions that relies on a small class of topological invariants in the line drawing of the object. Because these invariants are qualitative configurations of viewpoint-dependent features, becoming visible or occluded with changes in viewpoint relative to the object, the representation is a linked set of characteristic views defined by unique configurations of features (Freeman Chakravarty, 1980). The aspect graph approach has gained in popularity as computational methods for deriving aspect graphs from three-dimensional models have been developed (e.g., Eggert, 1991; Kriegman Ponce, 1990). Quite independently, there has been growing interest within psychology in the view-based approach to object representation. In particular, several researchers have demonstrated that object recognition of both novel and familiar objects is often viewpoint dependent (e.g., Bulthoff Edelman, 1992; Jolicoeur, 1985; Tarr Pinker, 1989). Such results led to the multiple-views hypothesis that objects are represented in human visual memory as a collection of viewpoint-specific images. In this approach, objects are recognized by normalizing an image of the perceived object to the nearest encoded view. One of the most crucial open questions in the multiple-views approach has been how such representations acquired and organized, and, specifically, what features are used within the representation and to delineate the boundaries between views.