Multi-GPU Image-based Visual Hull Rendering

Abstract

Many virtual mirror and telepresence applications require novel viewpoint synthesis with little latency to user motion. Image-based visual hull (IBVH) rendering is capable of rendering arbitrary views from segmented images without an explicit intermediate data representation, such as a mesh or a voxel grid. By computing depth images directly from the silhouette images, it usually outperforms indirect methods. GPU-hardware accelerated implementations exist, but due to the lack of an intermediate representation no multi-GPU parallel strategies and implementations are currently available. This paper suggests three ways to parallelize the IBVH-pipeline and maps them to the sorting classification that is often applied to conventional parallel rendering systems. In addition to sort-first parallelization, we suggest a novel sort-last formulation that regards cameras as scene objects. We enhance this method's performance by a block-based encoding of the rendering results. For interactive systems with hard real-time constraints, we combine the algorithm with a multi-frame rate (MFR) system. We suggest a combination of forward and backward image warping to improve the visual quality of the MFR rendering. We observed the runtime behavior of the suggested methods and assessed how their performance scales with respect to input and output resolutions and the number of GPUs. By using additional GPUs, we reduced rendering times by up to 60%. Multi-frame rate viewing can even be ten times faster.