3D model representation and manipulation based on skeletonization

Liang Shi
2011
3D model is a promising type of multimedia content for entertainment, research and education purposes. This thesis addresses the representation and manipulation of 3D models based on skeletonization, which is a commonly used technique to extract a compact descriptor and effectively capture the topological and geometric structure of 3D models. My research focuses on the refinement of skeletonization and its applications in 3D model matching, retrieval, and decomposition. By introducing a
more » ... k based on Scale-Space-Filtering (SSF), I integrate both the global node significance and the local chain-coded structure for pose-aware model retrieval; then adopt the topological mapping scheme for skeleton-based model decomposition. Experiment and comparison with state-of-art work on benchmark databases demonstrate the accuracy and efficiency of this framework. The first key contribution of this thesis is the improvement of skeletonization results. I introduce an adaptive skeletonization framework using SSF, and propose the notion of node significance and bending measurement to extract the structural features of 3D curve skeleton, hence deriving a more robust descriptor for model simplification and registration. By adopting SSF, I conduct dynamic skeleton pruning and smoothing based on the initial results of thinning. Results of improved skeletonization are shown compared to other thinning methods along with time performances of four recent approaches. This approach integrates the connectivity and the effectiveness in computation, thus it achieves the balance between representation ability and speed requirement. Another key contribution is the development on topology matching and chain coding techniques for measuring model similarity. To demonstrate the effectiveness in 3D model classification, I validate the chain code encoded skeletons using the Princeton Benchmark database, with particular emphasize on distinguishing different poses of similar models. Finally in this thesis I present an effective approach for model decomposition with enhanced semantics. By extracting robust skeleton and mapping model surface nodes to the decomposed skeleton branches, this method identifies the topology and geometry information of the 3D model. Thus it results in more meaningful segmentation components. Experiments on 2194 model demonstrate the advantage of this framework, comparing to three state-of-art approaches. According to animal anatomy, the proposed method keeps superio fidelity on four-leg animal models. Overall, my research in this thesis proposes a novel approach to integrate the global information as well as the local geometry structure of 3D models into the process of skeletonization. Applications on model retrieval and decomposition fully proved its effectiveness and accuracy.
doi:10.7939/r3n616 fatcat:qtwl5ommy5h73lotgsoixiytue