Shape from perspective: A rule-based approach

Abstract A system which is capable of inferring some of the structure of a 3-dimensional scene from a single perspective line drawing is discussed. Closed from equations for the inverse of the perspective transformation are used as modular inference engines. These inference engines use hypothesized spatial relationships between world entities to compute the unknown quantities such as distances between points, camera position, and focal length. The result of applying such modular computations is shown to be order independent and stable. In the absence of high-level models of objects in the scene, a hypothesize-and-test approach can be used to interpret the organization of the structures in the scene. The search space of hypotheses can be improved by utilizing known semantics of spatial relationships.