In this paper we present an algorithm for generating structural descriptions of stereo image pairs by perceptual grouping. In the first step, straight line segments are extracted in both images and the relations among them are investigated. In the second step, the relational graphs are constructed and connected subgraphs are detected. Using these subgraphs and similarities among them, a new graph, termed correspondence graph, is created. Applying graph theoretical approach, features in one image corresponding to features in the other image are found in the third step. The method was tested on a set of aerial stereo images of urban scenes.
Keywords: Perceptual grouping, stereo vision, feature extraction, structural stereopsis, scene description