Accelerated Bundle Adjustment in Multiple-View Reconstruction

Bundle adjustment is one of the main tools used for multiple view reconstruction. It seeks to refine the estimate of the 3D scene and the view parameter that minimize a certain cost function, e.g. the overall reprojection error. In this paper we propose a new algorithm to simplify the computation of the reprojection error for multiple views. With this algorithm, the bundle adjustment will be accelerated, whether the cameras are calibrated or un-calibrated. The proposed techniques for bundle adjustment are not only tolerant of missing data, but also allow the assignment of individual covariance to each image measurement. Experiments are conducted both on synthetic data and on real data to compare the proposed bundle adjustment techniques with the other existing methods. It is shown that the result obtained with the proposed techniques is comparable to that with the maximum likelihood estimation, however, it is more efficient.