Exploiting Redundancy for Large Scale Bundle Adjustment: In Partial Defense of Minimization by Alternation

Paper i proceeding, 2023

This paper presents a new approach to accelerate large-scale bundle adjustment (BA). As starting point, we empirically demonstrate that there exist a significant amount of redundant information that has yet to be leveraged in many real-world BA datasets. We propose an adaptive algorithm that builds on this redundancy and further utilizes the bipartite dependency structure of BA. Our algorithm maintains a small set of “training” 3D landmarks that is used to update the camera parameters, while the remaining landmarks are updated using faster point iterations (PI) in an inexact manner. This training set of landmarks is extended as necessary as indicated by a suitable “overfitting” criterion. The proposed algorithm also works gracefully for BA instances with robustified costs (such as the robust Cauchy and Geman-McClure costs). Experimental results on several large-scale BA datasets show that our method achieves faster convergence rates compared to several existing standard approaches.