Pose-Independent Object Representation by 2-D Views

We here describe a view-based system for the pose-independent representation of objects without making reference to 3-D models. Input to the system is a collection of pictures covering the viewing sphere with no pose information being provided. We merge pictures into a continuous pose-parameterized coverage of the viewing sphere. This can serve as a basis for pose-independent recognition and for the reconstruction of object aspects from arbitrary pose. Our data format for individual pictures has the form of graphs labeled with Gabor jets. The object representation is constructed in two steps. Local aspect representations are formed from clusters of similar views related by point correspondences. Principal component analysis (PCA) furnishes parameters that can be mapped onto pose angles. A global representation is constructed by merging these local aspects.