Using partial shape knowledge for stereo reconstruction

We describe a scheme for reliable and accurate surface reconstruction when the stereoscopic images contain only fine texture and no stable high level features. The approach uses partial shape information to improve the reliability of the surface computation process-by first fitting an approximate, global, parametric model, and then refining this model via local correspondence processes. A simple yet adequate parametric shape model is adopted, and the partial shape information is used to constrain the search for the optimal parameters. We show that even with very limited information regarding the surface shape (e.g., the surface is higher in the middle than on either side) a significant improvement is achieved, in the accuracy and reliability of the reconstruction