3D surface recognition and shape estimation from textured images

The authors present a novel and unifying decision-theoretic model-based approach for solving the problem of 3-D surface recognition and orientation estimation given an surface image patch. They concentrate on the subclass of quadric surfaces, mainly planes, cylinders, and spheres. The underlying assumption is that the different surfaces are textured. The authors adopt a model-based approach, wherein in the sensed image data that emanate from a 3-D surface patch are locally approximated by a homogeneous Markov random field (MRF) texture model which is parameterized by a few (known or unknown) parameters. To classify a surface image patch successfully, allowance is made for a direct link between the image data and the surface from which they emanate. This is achieved by explicitly making the 3-D surface shape parameters, the camera geometry, and the scene lighting part of the MRF model structure. Because of this link, optimum surface parameters are obtained, and minimum error bounds can be derived