Recognising rotationally symmetric surfaces from their outlines

Recognising a curved surface from its outline in a single view is a major open problem in computer vision. This paper shows techniques for recognising a significant class of surfaces from a single perspective view. The approach uses geometrical facts about bitangencies, creases, and inflections to compute descriptions of the surface's shape from its image outline. These descriptions are unaffected by the viewpoint or the camera parameters. We show, using images of real scenes, that these representations identify surfaces from their outline alone. This leads to fast and effective recognition of curved surfaces.