In this paper, we present an automatic approach to match correspondences on 3-D human bodies in various postures so that feature points can be automatically extracted. The feature points are very important to the establishment of volumetric parameterization around human bodies for the human-centered customization of soft-products (Trans. Autom. Sci. Eng., vol. 4, issue no. 1, pp. 11-21, 2007). For a given template human model with a set of predefined feature points, we first down-sample the input model into a set of sample points. Then, the corresponding points of these samples on the human model are identified by minimizing the distortion with the help of a series of transformations regardless of their differences in postures, scales or positions. The basic idea of our algorithm is to transform the template human body to the shape of the input model iteratively. To generate a bending invariant mapping, the initial correspondence/transformation is computed in a multidimensional scaling (MDS) embedding domain of 3-D human models, where the Euclidean distance between two samples on a 3-D model in the MDS domain corresponds to the geodesic distance between them in ℜ3 . As the posture change (i.e., the body bending) of a human model can be considered as approximately isometric in the intrinsic 3-D shape, the initial correspondences established in the MDS domain can greatly enhance the robustness of our approach in body bending. Once the correspondences between the surface samples on the template model and the input model are determined after iterative transformations, we have essentially found the corresponding feature points on the input model. Finally, the locations of the based local matching step.