Highly Parallel Algorithms for Visual Perception Guided Surface Remeshing

Visual perception can effectively capture visually salient regions of a 3D shape and guide the viewer's attention. In this thesis, we introduce the idea of 3D mesh saliency as a measure of regional geometry information provided by polygon meshes. We define the visual importance of a vertex based on its vertex curvature entropy which reflects the visual variation of the region centered at this vertex. We observe that such a definition of visual importance is capable of capturing visually interesting regions on a mesh. Incorporating our visual importance measure into the processing of 3D two-manifold meshes, more visually pleasing results are obtained compared to using a purely geometric measure of shape, such as curvature.;In order to be able to fast and efficiently process diversified models, a highly parallel remeshing framework based on meshfree techniques for processing surface sample points is also exploited. In this framework, the visual perception information is extracted in the image space and then mapped back to the Euclidean space. Based on these cues, a saliency field is generated to re-sample the input model. Lastly, a new projection operator named as AWLOP which extends Weighted Locally Optimal Projection (WLOP) to an adaptive one is developed to further optimize the distribution of re-sampled points. Experimental results demonstrate that our algorithm can remesh diverse polygonal models to well-shaped triangular meshes with high visual fidelity.;To benefit AWLOP, novel parallel algorithms for accelerating spherical range- search (SRS) for dynamic point sets that exploit the computational capabilities of current many-core GPUs are also investigated. Besides AWLOP, exact SRS is also required in geometry processing and physical simulation to avoid missing small features. We adopt an optimal AABB-tree as acceleration structure. With the help of a balanced AABB-tree, the spatial coherence of query points and the temporal coherence of dynamic points are exploited in our approach to achieve very efficient range-search and hierarchical update. We test our coherent SRS on a few applications including point-set geometry processing, distance-field generation and particle-based simulation. On a PC with NVIDIA GTX 660 Ti GPUs, our approach can take 1M queries on 1M dynamic points at a rate of 1600 queries/ms, where 49 neighbors are found in average within the range of 1/100 of the boundingbox's diagonal length. Moreover, we observe up to 5 times speedup over the voxelbased search running on GPUs.