Chinese Journal of Computers   Full Text
Title	GPU Based Real-Time Rendering of Spherical Depth Image
Authors	ZHU Jian1) WU En-Hua1),2)
Address	1)(Computer Graphics and Multimedia Laboratory, Department of Computer and Information Science, Faculty of Science and Technology, University of Macau, Macao) 2)(State Key Laboratory of Computer Science, Institute of Software, Chinese Academy of Sciences, Beijing 100190)
Year	2009
Issue	No.2(231—240)
Abstract &   Background	Abstract This paper presents an image-based real-time rendering algorithm on GPU. By the algorithm, we uniformly sample an object in all directions in a polar coordinates system, to construct a Spherical Depth Image. With two deduced Warping Equations, we then pre-warp the image onto a view-dependent tangent plane of the bounding sphere of the object to get an intermediate image, which can be further rendered onto the target image plane using standard texture mapping. By exploiting the inherent parallelism of modern programmable GPU, we transport the pre-warping process into Vertex Shader. Furthermore, the hardware pipeline’s rasterization function is utilized to conduct the image re-sampling efficiently to generate hole-free rendering results. Besides, Per-pixel lighting and Environment mapping effect is implemented in Pixel Shader to get high quality lighting effect. Finally, we solve the problem of limited field of view, and design a LOD (Level of Detail) algorithm to implement a real-time walkthrough system, in which all the rendered objects are occluded by each other correctly. Keywords GPU(Graphics Processing Unit); spherical depth image; image-based rendering; pre-warping Background Real time rendering of highly realistic scene has been a goal pursued by computer graphics researchers for decades. The parallelism of programmable hardware provides us a way to speed up, and image-based rendering techniques provide us a way to produce realistic images. Relief texture mapping (RTM) can efficiently represent the 3D surface details of objects, but its multi-pass pre-warp and interpolation for the hole-filling by software dominates the rendering time, which is computationally expensive and time-consuming. Moreover, In RTM an object representation consists of six relief textures associated with the cubic surfaces. This kind of representation has to be greatly over-sampled at the edges and the corners of the bounding box, which increases much unnecessary overheads. Fujita et al. use hardware-assisted per-pixel shader functions to accelerate relief texture mapping by offloading most of the original computations from CPU to GPU. However, it simply fills in the holes, caused by different sampling rate between source and target images, with nearby pixels; the quality of rendered images is usually poor. Donnelly implemented per-pixel displacement mapping in graphics hardware using 3D texture to speed up searching for the backward mapping method，but 3D texture requires such large volume of storage that it is difficult to represent large-scale scenes. Bar texture mapping uses lines as rendering primitives, but hardware rasterization is more efficient to polygons, in our method we use triangles for rasterization. Double Projective Cylindrical texture Mapping (DPCM) uses Double Projective Cylindrical Image (DPCI) for object representation, and puts the whole pre-warping process into Vertex Shader on GPU. As a result, its rendering speed is faster. However, DPCI is not a complete representation of the rendered object, there may be holes at the resultant images due to lack of source information. Even worse, DPCM has heavily limited field of view. To solve these problems, the authors proposed a GPU accelerated real-time depth image based rendering algorithm. The main contribution of this paper is as follows. The proposed Spherical Depth Image is sampled completely and uniformly, so the resultant image is complete without holes and the rendering overheads are minimized. By transporting most of the computation into GPU, we can greatly increase the rendering speed. Finally, we solve the problem of limited field of view. Experiment results show that it is suitable for massive real-time walkthrough system. This work is supported by National Key Fundamental Research and Development Project (973) with Grant No.2009CB320802, National High-tech R&D Program (863 Program) with Grant No.2008AA01Z301, National Natural Science Key Foundation of China with Grant No.60833007, the research grant and postgraduate studentship of University of Macau.