Researches On Generation And Compression Of Dynamic Mesh Sequence

The generation and compression of dynamic meshes are two fundamentals of computer animation. To generate dynamic meshes, there is a variety of ways including acquisition by 3D sensors, shape interpolation based on key frames, physical deformation, sequence editing and manipulation. Without need of too much interaction, shape interpolation is popular in animation design due to its easy and intuitive control. Sequence editing provides animators with flexibility to obtain more personalized dynamic mesh sequences. On the other hand, generated dynamic mesh sequences tend to take large room to store, which requires specific compression techniques for efficient storage. The topic of dynamic meshes compression draws attentions to achieving higher compression rate with lower visual distortion. This thesis studies interpolation based generation, editing, motion transfer and compression of dynamic mesh sequences.We propose a novel fast as-isometric-as-possible(AIAP) shape interpolation method. The shape interpolation problem is modeled as finding an AIAP motion trajectory from the start shape to the end shape, which leads to a minimization of a nonlinear energy. To initialize the nonlinear solver correctly, we propose a propagation-optimization algorithm using connection maps based on orthogonal local frames. The algorithm deals well with large-scale interpolation efficiently. In the phase of iterative refinement, we employ a block coordinate descent method that decoupled the original large linear system into many small-scale systems, each of which involves only one sequence of an edge, and then simplify the way of solving the systems. As for the operations of sequence editing and manipulation, we modify the relative velocity between any adjacent vertices to change poses across the given sequence. Finally, shape interpolation, editing and motion transfer are integrated into a uniform computational framework. Experiments show that the proposed algorithm of AIAP shape interpolation outperforms state-of-the-art methods both in quality and in efficiency.A new approach is proposed to compress dynamic meshes with the same connectivity.Noticing that the redundancy of information might increase when aligning all poses across the given sequence, we analyze and transform the trajectories of vertices of the given motion object, in order to find a subspace where the transformed trajectories correlate linearly as well as possible. We first perform a motion analysis to extract near-rigid parts of the object such that vertices of each part have almost the same rigid transformation during the motion. Second, to guild the alignment of rigid parts, we introduce a low-rank decomposition scheme that applies to the matrix composed by trajectories of vertices, and the transformed trajectories are reduced to ones with lower dimension via PCA. Finally, the rigid transformation, the bases and blending coefficients of PCA are further compressed by existing algorithms. Experiments reveal that our method achieves comparable compression rate comparing with the state-of-the-art approach,given the same visual distortion. In the case of dealing with nearly-rigid motion, our method has an edge in compression rate.