Contents I. Introduction
For 3DTV, several video formats have been developed such as the Left + Right stereo format and the Multi-view Video + Depth (MVD) format. One of the advantages of the latter format is that it enables rendering of new views by employing Depth Image Based Rendering (DIBR) for stereo or multi-view 3D displays [1]. When only one (Video + Depth)-stream is used, the quality of the synthesized views degrades rapidly with an increasing baseline [2]. Therefore, considerable efforts have been taken to improve view synthesis using DIBR with multi-view video. The main advantage is that artifacts occurring during DIBR from one camera, can be compensated by other cameras, resulting in a better rendering quality. Most recent view synthesis algorithms employ two nearest cameras for rendering an intermediate view [3] [4][5][6][7]. These algorithms have been designed primarily for a high rendering quality and not for a high efficiency. Moreover, the algorithms are mainly evaluated with a general-purpose CPU platform, making real-time view synthesis execution impossible. The possibility of accelerating DIBR algorithms using application-specific hardware, has been reported in literature. Bondarev et al. [8] explored an implementation of a free-viewpoint 3DTV receiver system on a FPGA board, but found that the current hardware provisioning is insufficient for elegant and efficient architectural solutions. Recently, graphics chips have become generic for computing (GPGPU) due to the easier programmability and increased generalization of GPU architectures [9]. Lee et al. [10] and Rogmans et al. [11] employ a graphics processor for accelerating DIBR algorithms by modifying DIBR data such that it executes in a graphics API pipeline. The use of a GPU for real-time DIBR is further explored in [12], where the author gives recommendations for fully exploiting the GPU resources, by combining the old graphics API (Direct3D, OpenGL) with the new general-purpose CUDA architecture. The previous proposals are straightforward system studies whereas in our case we aim at a real-time execution. With the ongoing development of better hardware and more efficient software, a real-time free-viewpoint 3DTV system using full 1080p HD video with off-the-shelf hardware comes within reach. For this purpose, we have implemented an efficient free-viewpoint DIBR rendering algorithm using a state-of-the-art GPU, since the GPU renderer forms a crucial component of such systems. As a basis for our GPU implementation, we have adopted the Free-Viewpoint Rendering algorithm (FVR) from [7]. Experiments have shown that due to highly parallel GPU programming, we have achieved real-time free-viewpoint DIBR operation with 1080p HD multi-view video.
