Accurate, Scalable and Parallel Structure from Motion

In this paper, we tackle the accurate Structure from Motion (SfM) problem, in particular camera registration, far exceeding the memory of a single compute node. Different from the previous methods which drastically simplify the parameters of SfM, we preserve as many cameras, tracks and their corresponding connectivity as possible for a highly consistent and accurate SfM. By means of a camera clustering algorithm, we divide all the cameras and associated images into clusters and leverage such formulation to process the subsequent track generation, local incremental SfM and final bundle adjustment in a scalable and parallel scheme. Taking the advantages of both incremental and global SfM methods, we apply the relative motions from local incremental SfM to the global motion averaging framework and obtain more accurate and robust global camera poses than the state-of-the-art methods. We intensively demonstrate the superior performance of our method on the benchmark, Internet and our own challenging city-scale data-sets.