GPU深度渲染：1280x720实时高质DIBR方法

130 浏览量更新于2024-08-30 收藏 251KB PDF 举报

本文主要探讨了在GPU上实现基于深度图像的实时渲染（Real-time Depth-Image-Based Rendering, DIBR）技术，针对1280×720分辨率的场景。作者提出了一种深度适应性预处理和超分辨率相结合的方法，旨在提升渲染质量和速度，充分利用GPU的并行计算能力，从而实现实时的深度图像渲染效果。首先，对于立体3D（Stereoscopic 3D, S3D）视频，它作为一种新兴媒体形式，通过模拟人眼视觉系统（Human Visual System, HVS）的工作原理，提供给观众更为真实和沉浸式的观影体验。S3D视频通常由双目立体成像系统捕捉，并在3D电视等设备上播放，创造出立体效果。文章的重点在于介绍了一种创新的GPU策略，即深度图像的深度适应性预处理。这种方法通过精细调整深度图像，确保在不同场景和光照条件下，渲染出清晰、准确的深度信息，从而提高渲染的细节表现力。此外，文章还提到了超分辨率技术的应用，通过提升图像分辨率，进一步增强画面的细腻度和清晰度，使深度图像在渲染时呈现出更高的视觉质量。为了实现真正的实时性能，作者强调了GPU的并行计算优势。GPU架构特别适合大规模并行处理任务，如深度图像的计算密集型操作。通过将深度图像渲染分解为多个并行执行的任务，可以显著缩短处理时间，确保整个过程能在合理的时间内完成，满足实时渲染的需求。实验结果显示，与现有方法相比，作者提出的深度图像基于GPU的实时渲染方法在渲染质量和计算效率上都表现出色。关键词包括深度图像渲染（DIBR）、GPU、洞填充（hole filling，可能指的是对深度图中的空洞或缺失部分进行修复）、并行计算以及实时性，这些是理解该研究的核心要素。总结来说，这篇文章深入探讨了如何利用GPU的强大性能和先进的图像处理技术来优化深度图像渲染，以提供高质量且实时的立体3D视觉体验。这对于视频编码、游戏开发以及虚拟现实等领域具有重要的实际应用价值。

Real-time Depth-Image-Based Rendering on GPU

Zengzeng Sun

SRC-Xian DMC Lab/SSG, Samsung Electronics

Xi’an, China

zengzeng.sun@samsung.com

Cheolkon Jung

School of Electronic Engineering, Xidian University

Xi’an 710071, China

zhengzk@xidian.edu.cn

Abstract—In this paper, we propose real-time depth-image-

based rendering (DIBR) on GPU for 1280×720 resolution. We

utilize depth adaptive preprocessing and super-resolution to

achieve high-quality DIBR. Moreover, we employ GPU-based

parallel computing to achieve real-time DIBR. Experimental

results demonstrate that the proposed method achieves

superior performance in comparison with the existing methods

with respect to rendering quality and computing time.

Keywords- Depth-image-based rendering (DIBR), GPU, hole

filling, parallel computing, real-time.

I. INTRODUCTION

Stereoscopic 3D (S3D) videos are considered as a new

media which provides a realistic and immersive experience

to viewers [1]. They are captured by stereo imaging systems

imitating the human visual system (HVS), and played in S3D

displays such as 3DTV which provides viewers with 3D

effects [2]. Because they are recorded by 3D stereoscopic

cameras like 2D situations, there are some critical issues

such as synchronizing the stereoscopic cameras [3], storing

and transmission of a huge amount of data [4], etc. In

practice, future S3D displays provide viewers with the

services of navigating through the scenes to choose an

appropriate viewpoint depending on their positions for

reality and immersive, which is known as free-viewpoint

television (FTV) [5]. It is of more practical use to use S3D

videos using stereoscopic cameras. The existing 2D videos

should be converted into 3D ones to play on S3D displays.

Thus, recent studies have focused on virtual view synthesis

which generates various intermediate views using limited

number of camera images. Depth-image-based rendering

(DIBR) [6] is a well-known technique of virtual view

synthesis which generates novel views from a 2D reference

image and its corresponding depth image. A coding standard

for the video-plus-depth representation, known as MPEG-C

part 3 [7], is issued by MPEG due to the compression-

efficiency for transmission and compatibility with the

existing 2D broadcasting system. DIBR is also a core

technology in advanced three dimensional television system

(ATTEST 3DTV System) [8] whose goals include a 2D

compatible coding and transmission scheme for S3D videos

using MPEG-2/4/7 and a flexible and commercially feasible

3DTV system for broadcast environments. For the virtual

view synthesis, the 3D warping should be performed to map

This work was supported by the National Natural Science Foundation of

China (No. 61271298) and the International S&T Cooperation Program of

China (No. 2014DFG12780).

the pixels in the 2D reference image into desired virtual view

using explicit geometric information, depth maps, and

camera viewpoint parameters [9]. However, the 3D warping

generates obvious holes in the synthesized virtual view

mostly caused by the reason that occluded areas in the

reference image become visible in the synthesized virtual

view. The hole problem is an inherent problem in DIBR

because it has a great influence on the visual perceptual

quality of the generated S3D videos.

Various techniques have been proposed for hole-filling

[10]-[18]. They are classified into two main groups:

interpolation based hole-filling [10]-[14] and inpainting-

based hole-filling [15]-[17]. Recently, Solh et al. have

proposed a novel hole-filling approach in DIBR referred to

as hierarchal hole-filling (HHF) that fills in the holes using a

pyramid-like structure [18]. HHF yields a better rendering

quality than the other approaches, but suffers from blurred

results in the hole area around the object because the

upsampling from lower resolution produces blurred effects.

It is noteworthy that Zhang et al. [11] first introduce depth

map preprocessing into DIBR framework (see Fig. 1) which

is composed of three major components because a depth map

with horizontal sharp edges would result in big holes after

3D warping. They reduce the size of big holes before 3D

warping by smoothing the edges of the depth sharp transition,

which can make the subsequent hole-filling much more

easier. However, depth map preprocessing may bring about

the problems of geometric distortions and losses in depth

cues [19]. Thus, several approaches have been proposed to

reduce the artifacts of depth map preprocessing, but still

suffer from the annoying phenomena and remarkably

increasing the computational complexity. To implement

image processing algorithms in fast speed or real-time, the

parallelization based on graphics processing unit (GPU) have

recently become an active topic in the community [20], [21].

Real-time processing is required in the practical applications,

and the resultant quality of implementation should be

ensured under the real-time constraint as well.

Fig. 1 DIBR with depth map preprocessing, redrawn from [11].

2015 International Conference on Cyber-Enabled Distributed Computing and Knowledge Discovery

DOI 10.1109/CyberC.2015.97

324

下载后可阅读完整内容，剩余4页未读，立即下载

weixin_38620839

粉丝: 8
资源: 938

GPU深度渲染：1280x720实时高质DIBR方法

GPU加速的超声波图像实时重建Matlab程序

GPU实时渲染大涡模拟技术研究

GPU Pro 5深度解析：高级渲染技术

基于GPU的实时深度图像前向映射绘制算法.pdf

基于GPU加速的深度图像绘制.pdf

基于GPU的球面深度图实时绘制.pdf

基于GPU的多层显微图像实时融合技术.pdf

基于GPU编程的三维烟雾实时渲染.pdf

数字地球上基于GPU的海水渲染方法.pdf

基于GPU的近似软影实时绘制.pdf

最新资源