Rapid invocation and matching of 3D models to 2D images using curvilinear data

In 2D images of 3D objects it is not possible to account for 3D viewpoint invariance as those invariant relations cannot be found. Therefore, a process of rapid hypothesis and verification is required in which the perceptual groupings of curvilinear features provide the hypotheses for verification by back-projection of the model data. In this paper, we present one approach to this latter cycle, using model representations having both straight and curved elements. We have described a practical approach to match 2D scenes to 3D models in which the scene data is described as a set of curvilinear features extracted from a grey scale image of a rigid object taken from an arbitrary viewpoint, and the model data is derived from a CSG modeller.