PhD Thesis: 3D Free-form Surface Representationand Its Applications [pdf]
Yi Song, Li Bai
The increasing availability of reliable three-dimensional (3D) sensors and sensing algorithms has boosted research in 3D face recognition. This project aimed to find a sufficient 3D shape representation amenable to surface matching, where it is 1) precise enough to distiguish intra-class differences and, 2) computationally feasible for one-to-many matching against large face databases. Therefore a novel 3D modelling technique is developed to automatically reconstruct B-Spline surfaces from point clouds (i.e. unstructured data) directly (Fig. 1), from which the dense correspondences between objects via a common parameter space can be automatically established.

un-textured rendering results. texured rendering results.
Figure 1 B-Spline surface reconstruction from points cloud. The reconstructed face model is presented by 616 shape descriptors while polygon model is composed of 16,247 triangles.
In addition to face recognition, this technique can be used to model other more complex 3D shapes. Figure 2 shows one patch B-Spline surface modelling of armadillo's right leg. Figure 3 illustrated female upper body reconstruction by using four-patch B-Spline surface.
Figure 2 Armadillo's right leg modelling.
Figure 3 Female upper body reconstruction.
One of the advantages of our B-Spline reconstruction technique is its ability to discretise the continous surface representation in any required resolution, as shown in Figure 4.

Figure 4 Three meshes extracted at different resolution (left). The right figure is the same object rendered at different level of detail (LOV).
As a canonical mapping from the reconstructed surface to a unit plane always exists, the dense correspondences between objects can be established automatically. Thus this modelling technique is easily extended to the field of 3D metamorphosis, as shown in Figure 5. More examples are demonstrated in Gallery.

Figure 5 Automatically generated smooth 3D metamorphosis sequences.