Scene modeling: a structural basis for image description

Abstract Conventional statistical approaches to image modeling are fundamentally limited because they take no account of the underlying physical structure of the scene nor of the image formation process. The image features being modeled are frequently artifacts of viewpoint and illumination that have no intrinsic significance for higher-level interpretation. This paper argues for a structural approach to modeling that explicitly relates image appearance to the scene characteristics from which it arose. After establishing the necessity for structural modeling in image analysis, a specific representation for scene structure is proposed and then a possible computational paradigm for recovering this description from an image is described.