Research On Acceleration Techniques For Real-time Rendering Of Highly Detailed Surfaces

Real world objects, such as sculptures, archeological artifacts and even natural objects, often require highly detailed representations. The highly detailed surfaces can be obtained using 3D scanner, and thus make large volume of datasets. Real time visualization of highly detailed surfaces is a challenging research task.Modern Graphics Processing Unit (or GPU) has powerful parallel computing capability, and thus using GPU to accelerate the real time rendering of highly detailed surfaces is an effective approach. However, GPU has its distinct features different from CPU in architecture and programming. This dissertation tries to develop new data structure and algorithms for accelerating the LOD selection and detail lighting through GPU, and implement the interactive transmission of the highly detailed surfaces. The main contributions of the dissertation are summarized as follows:Firstly, the dissertation proposes a general GPU-accelerated LOD framework. The framework is based on a quadtree structure, called P-Quadtree, which is constructed from the geometry image (or GIM), the image-like representation of 3D surfaces. This allows geometry and the LOD model to flow more naturally as a signal through the current GPU rendering pipeline to accelerate the realistic light computation, multi-resolution processing and digital geometry processing (DGP). To stitch the boundary of cutting charts effectively, the approaches to the newly extended geometry images and seamless geometry image are proposed. P-Quadtree is served as the infrastructure of the works in the dissertation.Secondly, a novel algorithm that is fully implemented on GPU to accelerate dynamic LOD is proposed. To perform LOD selection on GPU efficiently, the hierarchical structure of P-Quadtree is converted into LOD texture atlas to fit the architecture of GPU. The approach includes two rendering passes: the LOD selectionthat is performed in the fragment shader, and the culling and triangulation operations that are performed in the vertex shader. The proposed algorithm improves the efficiency of LOD selection by utilizing the parallel computing power of GPU, and the load of CPU is alleviated.Thirdly, a GPU-accelerated multi-resolution rendering approach is proposed. The realism of the simplified meshes can be enhanced greatly by normal mapping, which can be accelerated by per-pixel lighting technique furnished by modern GPU, but the silhouette appearance is poor. Based on the characteristics in human vision, the proposed approach employs view-dependent LOD based on P-Quadtree to simplify the meshes. Then geometry complexity of the meshes is reduced effectively, and meanwhile, the rendering realism can be enhanced and the shape on the silhouette can be preserved. Hybrid rendering of polygon and point can be utilized to accelerate the image synthesis.Lastly, an interactive transmission scheme over network is proposed to transmit the geometry and detail of the model progressively. An efficient compression technique is employed to encode the hierarchical structure and vertices of its nodes by utilizing the regular structure of P-Quadtree. By taking advantage of normal texture atlas, the rendering result of the partial transmitted model is improved greatly. Since the flexibility of the nodes transmission could be achieved by using the view-dependent LOD techniques, only the nodes on the silhouette to the current viewpoint need to be refined and transmitted. This leads to great reduction of the amount of data needed to transfer each frame during network navigation.