Block-column iterative bundle adjustment for large-scale 3D reconstruction

Abstract

Large-scale Bundle Adjustment (BA) poses a significant computational challenge in 3D reconstruction and Structure from Motion (SfM). Traditional BA solvers, such as the Levenberg-Marquardt (LM) and Dogleg (DL) algorithms, become increasingly time-consuming as the scale of the Reduced Camera System (RCS) grows. In this work, we present a novel algorithm designed to address the complexities of large-scale BA. Our method utilizes block-column iterations to exploit the problem’s sparse structure, improving solver efficiency. Experimental validation on the Bundle Adjustment in the Large (BAL) dataset shows that our approach outperforms conventional iterative methods, significantly accelerating the solution of the normal equations and speeding up large-scale 3D reconstruction. Furthermore, our algorithm can be seamlessly integrated as a sub-problem solver within the Stochastic Bundle Adjustment (STBA) framework, enhancing both the speed and accuracy of distributed optimization.