Implementation of fast free-viewpoint video rendering on graphics processing units

Abstract

This paper presents the development of a fast Free-Viewpoint Video (FW) rendering algorithm that exploits the parallelism offered by General Purpose Graphics Processing Units (GPGPUs). The system generates virtual views through the use of Depth Image-Based Rendering (DIBR) algorithms, implemented using NVidia® Compute Unified Device Architecture (CUDA). A novel reference image brightness adjustment algorithm that exploits the correspondences between matching pixels in the reference images to avoid drastic brightness switching while navigating in between views is also discussed. The developed solution ensures that data transfers are kept at a minimum, thus improving the overall rendering speed. Objective and subjective test results show that, for typical free-view scenarios, the proposed algorithm can be successfully deployed in real-time FW systems, providing a good Quality of Experience (QoE).