Space-Time Representation and Editing of 3D Video Mesh Sequences.

Advances in surface performance capture have enabled the reconstruction of real world scenes such as people and animals with a realism hard to achieve by an animator. The work presented in this thesis aims to develop techniques for interactive editing and manipulation of captured mesh sequences with the flexibility associated with conventional computer animation techniques. In particular, the application of Laplacian deformation for animation and compression of surface motion capture data is investigated. Laplacian deformation enables the manipulation of a mesh at a vertex level while maintaining its local geometric properties but lacks a mechanism for ensuring the preservation of its underlying physical structure. Motivated by this limitation, a learnt surface deformation basis constructed in the space of differential coordinates is introduced. The incorporation of this basis into the Laplacian framework constrains the solution to the space of plausible deformations built from a set of examples, therefore preserving the structure of the mesh. The successful application of this approach to space-time editing together with a set of novel non-linear edit propagation techniques are presented. Representations for efficient storage of surface motion capture sequences, generally comprised of hundreds of frames with thousands of vertices, are investigated. A novel layered representation that exploits the articulated nature of the data is presented and compared with other compression techniques based on PCA and Laplacian deformation, with and without using the aforementioned surface deformation basis. The proposed layered representation achieves consistently high compression ratio with low maximum reconstruction errors in three test sequences from different characters. Learning full body models of surface deformation presents the drawback of encoding correlations between body parts, which limits the editing flexibility. Independent part-based surface deformation models and hierarchical propagation of constraints are presented as a solution for this limitation. Pose editing performed on three different characters demonstrate that this technique broadens the range of plausible poses achieved using learnt Laplacian deformation. Furthermore, comparison with traditional skinning methods illustrates how this method synthesises more natural poses by reproduction of the non-rigid deformation observed in the example meshes