Conformation-based Parameterization of Scanned Face Data

    Nowadays, a variety of methodologies developed for 3D shape capture are available, ranging from laser range scanning (e.g., CyberWare, VIVID), and stereo photogrammetry (e.g., EyeTronics, 3QTech), to structured light projection. With these devices or techniques, one can capture and digitize the complete surfaces of a large number of human faces accurately. For some applications, such as face shape analysis, we are more concerned with applying geometry processing algorithms to a group of models than to a single one. However, when given a group of scanned shapes, to perform simple operations, such as computing their average shape or computing the norm of their differences, is not trivial. The main cause for this difficulty is the generally different sampling patterns used to describe the geometry. Geometry processing algorithms involving multiple models require a consistent parameterization and a common sampling pattern. A dense set of correspondences between the individual instances of the collection of scanned shapes need to be established. Moreover, the scanned data typically consists of hundreds of thousands of 3D points. Such a dense data-set cannot be easily used for animation. Although mesh decimation results in a usable approximation of the scanned data, there is not enough control over the connectivity of the mesh. This can lead to artifacts in the animation due to misalignment of the triangle edges.
    In this project, we present a mesh conformation method for establishing a point-to-point correspondence among scanned 3D face shapes. Our method is based on fitting a deformable generic mesh which has a predefined configuration onto detailed human face range scans in a global-to-local fashion. The mesh conformation globally maps the generic mesh to the scanned 3D shape based on semi-automatically specified corresponding feature points through the application of a Radial Basis Function (RBF)-based volume warping. In order to generate more correspondences for a good match and to ease the task of manually specifying the correspondences, we develop an automatic procedure to refine the feature point sets. After global warping, a local deformation is carried out to fit all vertices on the generic mesh to the scanned surface. We formulate an optimization problem to solve for the local deformation using an energy function that is a combination of two measures: the proximity of transformed vertices to the scanned shape and the smoothness over the surface. We demonstrate reconstruction and parameterization of 186 human face scans in a large database. With our method, we explore a variety of applications, including texture transfer, morphing, and statistical analysis of shape. We have implemented our method in the context of modeling human faces; however, given the availability of a suitable generic model, the general techniques that we have developed could be applied to other objects.
    The three-step mesh conformation method: