In object recognition accurate segmentation of a particular object of interest (OOI) is critical. The OOI usually consists of a set of homogeneous regions with spatial relations among them. Thus, class-specific knowledge on the visual appearance and spatial arrangement of the regions can be useful in discriminating among objects from different classes. In this paper, we propose the use of the Attributed Relational Graph (ARG)-based formalism as a means of representing both visual and spatial information in a single structure. In the proposed framework, a training set of images, each of which contains an instance of the OOI, is given. Afterwards, each image is over-segmented into a set of visually homogeneous regions and the corresponding ARG is constructed. Given such graph representations, OOI model learning reduces to a subgraph matching problem.
展开▼