基于云计算的海量图片集场景D重构研究---毕业论文.docx

本 科 毕 业 论 文基于云计算的海量图片集场景3D重构研究Research on the Scene 3D Reconstruction from Large Image Collection based on Cloud Computing姓 名： 学 号：学　　 院：软件学院系：软件工程专 业：软件工程年 级：校内指导教师：　 年 月摘 要场景重构是指从无序图片集中自动生成三维场景几何的一项技术，尤其指拍摄照片的照相机的位置和方向的重构，以及场景中的3D坐标点的重构。然而，当图片集内的图片数量增多的时候（例如，由1,0000张图片构成的网络图片集），由于算法时间复杂度的原因，场景重构所需要的时间变得难以接受。在本论文中，我们将描述了一个新型场景重构系统，它首次将云计算技术应用于海量图片集的3D重构。借助于云平台，该系统大幅度缩短了场景重构的时间，并为场景重构过程增加了新特性（例如容错和数据备份）。为了能够将场景重构程序成功移植到著名的开源云平台Hadoop，我们首先将重构管道划分为4个MapReduce模块以实现并行计算，重新设计每一个模块的I/O操作并扩展了Hadoop的API来处理非ASCII文本的文件。实验结果表明，在一个由7台计算机节点构成的Hadoop集群上，我们设计的系统完成场景重构的时间比原系统快了大约7倍。关键词：场景重构；云计算；海量图片集AbstractScene reconstruction is a technique that takes an unordered collection of images to produce 3D scene geometry. In particular, the position and direction of camera for each photo and the 3D coordinates of points in the scene are recovered. However, when the image collection becomes large (e.g., an Internet image collection of 10,000 images), the wall clock time of scene reconstruction becomes unacceptable due to the high algorithm complexity of it. In this thesis, we describe an innovative system which first time applies cloud computation to large image collections of 3D reconstruction. With the great benefit of clod computation, our system reduces the time cost of scene reconstruction significantly and adds some new properties (e.g., fault tolerance, data back-up) to it. To successfully transplant scene reconstruction upon the famous open source cloud computing framework, Hadoop, we divide the scene reconstruction pipeline into 4 main MapReduce modules to implement parallel computing, redesign I/O operations of each module and extent the Hadoop API to deal with multiple file formats beyond ASCII text. By mean of evaluation, we show that our system performs about 7 times faster than the original system upon a Hadoop cluster of 7 machines.Key words: Scene reconstruction; Cloud computing; Large image collections目录第一章绪论11.1背景介绍11.2基于网络照片集的场景3D重构21.3场景重构所面临的挑战31.4云计算平台4第二章场景3