Representation of scenes from a multi-view perspective has recently attracted lots of research interest in the fields of computer graphics and computer vision, due to its effectiveness and efficiency. Utilizing the data coherence of multi-view representations can provide critical improvements in algorithm performance. This thesis is devoted to addressing problems in efficient rendering (including triangle reordering and stereo rendering), and 3D reconstruction, all of which exploit the data coherence of multiple viewpoints.

Two novel approaches are proposed for triangle mesh reordering, one for static models with tight front-to-back triangle orders and another for animated models with loose front-to-back orders. Briefly, these approaches precompute a small number of triangle o...[ Read more ]

Representation of scenes from a multi-view perspective has recently attracted lots of research interest in the fields of computer graphics and computer vision, due to its effectiveness and efficiency. Utilizing the data coherence of multi-view representations can provide critical improvements in algorithm performance. This thesis is devoted to addressing problems in efficient rendering (including triangle reordering and stereo rendering), and 3D reconstruction, all of which exploit the data coherence of multiple viewpoints.

Two novel approaches are proposed for triangle mesh reordering, one for static models with tight front-to-back triangle orders and another for animated models with loose front-to-back orders. Briefly, these approaches precompute a small number of triangle orders, each serving a cluster of viewpoints, thus eliminating reordering computation during runtime rendering. Compared to other algorithms that perform precomputing strategy, our proposed methods achieve state-of-the-art performance in terms of memory usage and rendering speed. The third work is concerned with efficient temporal and stereo reprojection using simplified meshes. The proposed approach improves reverse reprojection by approximating detailed rendering with a low-resolution model, thus reducing vertex processing when using a detailed model. Finally, we give a novel solution to 3D reconstruction from a set of views. In contrast to existing cost volume approaches, our method directly processes the target scene as point clouds. Taking advantage of the flexibility of point clouds, our method produces better reconstruction quality with less memory usage than state-of-the-art methods.

Overall, under the guiding principle of utilizing data coherence of the multi-view representation, my efforts have led to top-performing algorithms for efficient rendering and 3D reconstruction.