UC San Diego

Our vision is to enable efficient acquisition and synthesis of highly detailed 3D surface models that are also easy to animate in a plausible and realistic way. The state-of-the-art surface acquisition technology is range scanning, which can measure surface geometry with a high degree of accuracy and speed. However, the output is only a partial view of the surface that has much missing data, and there is no tracking of the surface motion in the case of a moving subject. To reconstruct a complete model of the subject, we must align multiple range scans taken from different times and viewpoints to fill in the missing data and track the motion of the surface. At the same time, we would like to fit a reduced deformable model that expresses the surface motion in terms of a few intuitive parameters. In this dissertation, we develop algorithms to process and align multiple range scans of a moving articulated subject. Our algorithms can automatically align multiple scans to a common pose, thus reconstructing the full geometry of an articulated subject along with a model of its motion. Our methods perform this alignment in a completely unsupervised way: without markers, a template, or a user-defined segmentation of the surface. A key contribution is the use of discrete optimization techniques to automatically estimate the articulated structure of the surface based on its motion. First, we describe a method to align a pair of 3D surfaces that is robust to large motions and much missing data. This algorithm samples rigid transformations between the surfaces and performs an alignment by optimizing an assignment of the transformations to the surface. Its robustness to large motions makes it useful for initializing a registration. Next, we present a technique to automatically fit an articulated surface motion model to a pair of range scans. We efficiently solve for the transformations and weights of this model by repeatedly estimating them in alternating fashion. The key benefit of this approach is that the solved model parameters can be used to easily and intuitively edit the pose of the scanned geometry. Finally, we improve and combine these two approaches to automatically reconstruct an articulated 3D model from multiple range scans. We reduce alignment error by simultaneously solving for the alignment of all input scans. We demonstrate that this method can reconstruct a variety of poseable, articulated 3D models from partial surface data acquired by a range scanner