Learning and recognition of 3D objects from appearance

Abstract

The authors address the problem of automatically learning object models for recognition and pose estimation. In contrast to the traditional approach, they formulate the recognition problem as one of matching visual appearance rather than shape. The appearance of an object in a two-dimensional image depends on its shape, reflectance properties, pose in the scene, and the illumination conditions. While shape and reflectance are intrinsic properties of an object and are constant, pose and illumination vary from scene to scene. They present a new compact representation of object appearance that is parameterized by pose and illumination. They have conducted experiments using several objects with complex appearance characteristics.<>