The choice of the most suitable generic shape model is an open problem in computer vision. Modelling provides a compact description of the raw data and provides parameters that allow tasks such as recognition and pose estimation to be performed. However most model driven methods are based on knowledge or expectations on how the data is structured. Such approaches do not fare well in unstructured environments since no a priori assumptions can be made of the environment. The approach presented here attempts to address this problem by building complementary descriptions of the data using both global and local models. Superquadrics and surface patches were chosen as the local modelling primitives. They respectively allow a model to express both gross qualitative shape and fine structural details. This approach to modelling is illustrated experimentally using laser range data. A superquadric is fitted to this data using a nonlinear least squares minimization (Levenberg-Marquardt) of the superquadric inside-outside function. The superquadric that results in the best possible fit is expressed in terms of its position, size, shape and pose parameters. The residual of the fit is then modelled at several scales using multiple surface patches with uniform mean and Gaussian curvature. A hierarchical ranking of these patches is used to describe the residual based on geometric properties. These geometric properties are ranked according to criteria expressing their stability and utility. The most stable patches are selected as the description of the residual. Using these two models, a composite representation in terms of a superquadric combined with a multiscale surface is generated. This representation can then be used for both pose estimation and object recognition.
N. Ayoung-Chee, G. Dudek, F. Ferrie