Recognition of polyhedral objects-aspect graph generation based on a learning-by-showing approach

A new method is proposed for the automatic generation of aspect graphs for modeling polyhedral objects. It follows the learning-by-showing approach, and does not require expensive computations based on a prior knowledge such as CAD models. The image representations are sets of visible vertices of an object on image planes, and given such vertex images taken from a number of viewpoints, the method classifies the images into different aspects by inferring topological homogeneity among them. The topological homogeneity is determined as results of vertex registration between any two vertex images. A viewpoint selection method based on spherical Voronoi diagrams is also developed for the purpose of viewpoint planning.