Three-dimensional object recognition is viewpoint dependent.

Abstract

The human visual system is faced with the computationally difficult problem of achieving object constancy: identifying three-dimensional (3D) objects via two-dimensional (2D) retinal images that may be altered when the same object is seen from different viewpoints1. A widely accepted class of theories holds that we first reconstruct a description of the object's 3D structure from the retinal image, then match this representation to a remembered structural description. If the same structural description is reconstructed from every possible view of an object, object constancy will be obtained. For example, in Biederman's2 oft-cited recognition-by-components (RBC) theory, structural descriptions are composed of sets of simple 3D volumes called geons (Fig. 1), along with the spatial relations in which the geons are placed. Thus a mug is represented in RBC as a noodle attached to the side of a cylinder, and a suitcase as a noodle attached to the top of a brick. The attraction of geons is that, unlike more complex objects, they possess a small set of defining properties that appear in their 2D projections when viewed from almost any position (e.g., all three views of the brick in Fig. 1 include a straight main axis, parallel edges, and a straight cross section). According to the RBC theory, a complex object can therefore be recognized from its constituent geons, which can themselves be recognized from any viewpoint.