EIGENSPACES FOR GRAPHS

In this paper, we investigate the feasibility of using graph-based descriptions to learn the view structure of 3D objects. The graphs used in our study are constructed from the Delaunay triangulations of corner features. The investigation is divided into two parts. We commence by considering how relational structures can be encoded in a way which can be used to generate parametric eigenspaces. Here we investigate four different relational representations derived from the graphs. The first three of these are vector encodings of the adjacency graph, the weighted adjacency graph, and the point proximity matrix; the fourth representation is the edge weight histogram. We study the eigenspaces which result from these different representations. In addition, we investigate how multidimensional scaling may be used to generate eigenspaces from a set of pairwise distances between graphs.