Hardware Architecture for Real-Time Depth-image-based Rendering System

This paper presents a hardware architecture for real-time depth-image-based rendering (DIBR) system. The shift-sensor camera setup is used in this system so that hardware-efficient algorithms can be applied to reduce the computational complexity and hardware cost. For the computational complexity, we propose a 3D image warping algorithm that requires no camera parameters to replace the complicated homographic transform, and a hole-filling method based on the disparity map to replace the complicated inpainting algorithm. For the hardware cost, we propose a fold-elimination approach based on view-judgment to replace the Z-buffer algorithm, and the row-level pipelining to reduce the requirement of external memory for disparity maps and reference images. Experimental results show that our implementation can achieve the similar synthesis quality as the quality implemented by software.