Abstract Bundle adjustment with very large scale datasets has drew much concern recently in both photogrammetry and computer vision communities. Different from the existing out-of-core and distributed methods for large scale datasets, we propose a fast and accurate bundle adjustment method which still uses the framework of the traditional Levenberg Marquardt (LM) algorithm while adopting preconditioned conjugate gradient (PCG) to iteratively solve normal equation, and using point resampling scheme and normal matrix compression to decrease the memory requirement and computational complexity. Preliminary results show that our method running on a single laptop computer with i7 2.6 GHz CPU and 8 GB RAM is even faster than the state-of-the-art distributed method deployed on a large distributed computer system with multiple computers each of which is equipped with CPU i7-4770K 3.5 GHz with 8 threads and 32 GB RAM and connected with each other at the speed of 10 MB/s. The proposed method is also more accurate.

中文翻译：

---

一种快速准确的超大规模数据束平差方法

摘要 最近，摄影测量和计算机视觉社区对超大规模数据集的捆绑调整引起了广泛关注。与现有的大规模数据集的核外和分布式方法不同，我们提出了一种快速准确的束调整方法，该方法仍然使用传统Levenberg Marquardt（LM）算法的框架，同时采用预条件共轭梯度（PCG）进行迭代求解正规方程，并使用点重采样方案和正规矩阵压缩来降低内存需求和计算复杂度。初步结果表明，我们的方法在带有 i7 2.0 的单台笔记本电脑上运行。6 GHz CPU 和 8 GB RAM 甚至比部署在具有多台计算机的大型分布式计算机系统上的最先进的分布式方法还要快，每台计算机配备 CPU i7-4770K 3.5 GHz 8 线程和 32 GB RAM并以 10 MB/s 的速度相互连接。所提出的方法也更准确。