We extend the original incremental SfM approach into a divide-and-conquer approach. Our approach is much more efficient than state-of-the-art open-source SfM systems (COLMAP, OpenMVG, TheiaSfM), while surpass the accuracy of them at the same time. The pipeline of our GSfM is shown below:
Thanks to our adaptive graph cluster algorithm, the images are divided into different groups. The images with strong connections are divided in the same group. And strong/weak MST conditions are used to enhance the connections between different clusters. After that, a robust incremental SfM approach (Based on an early version of OpenMVG) is performed in each cluster. While different clusters located in different reference systems, our graph-based merging algorithm is designed to automatically align the point clouds efficiently. Thus, our SfM approach is named GSfM.
