3-D shape recovery using a relaxation algorithm based on surface feature consistency

3D high-level surface features are widely used for recognizing complex curved objects. The features, however, are difficult to extract accurately from raw range images because the images are usually noisy for the required differential operations. This paper proposes a new 3D shape recovery method that obtains refined shape features by deriving an optimal approximation to the true surfaces in a raw image. In principle, the method is mainly based on a relaxation algorithm that reconstructs original shapes by means of feature consistency constraints. The application of these constraints makes the algorithm not only invariant with the object poses but also very adaptive to complicated shape changes. Meanwhile, surface discontinuities are preserved by utilizing a possibility function of discontinuity (PFD) in controlling the constraint propagation to prevent edge points from overslurred.