Partial bundle adjustment for accurate three-dimensional reconstruction

Abstract

Structure from motion aims at estimating the cameras and sparse structure of the scene globally, but the approximation biases in bundle adjustment and the propagated errors among the registered cameras and triangulated points reduce the quality of the reconstruction. To improve the accuracy of the reconstruction, the authors present an efficient subgraph-based approach to optimise the visibility graph locally. The central idea is to divide the visibility graph into well-configuration subgraphs for partial bundle adjustment based on a Markov random field, reconstruct the dense local points via the subgraph-based multi-view stereo matching, and then merge them into a consistent scene to extract the complete surface. They evaluate the proposed algorithm on several real and synthetic datasets, and the results demonstrate that their algorithm can improve the reconstruction quality significantly and maintain an inexpensive computation cost in the meanwhile.