Robust and Globally-Consistent 3D Reconstruction for Acquisition and Registration

3D reconstructions from photographs often exhibit poor robustness and global consistency. This is a problem in applications such as model acquisition and registration where global consistency is needed to ensure that the reconstruction looks acceptable from all views. Our goal is to develop a robust reconstruction approach which leads to better consistency between viewpoints. We do this by solving the 3D reconstruction problem in a way that does not involve the camera orientation. This produces a better method because the confusion between camera rotation and camera translation is known to be a source of problems. Our novel formulation consists of low-order polynomial equations involving only the camera centers and the observed scene points. The formulation significantly improves robustness and accuracy, despite significant noise in the measurements, and calculates a more globally-consistent solution that reduces pixel-reprojection errors for views near and far from captured viewpoints, produces a robust refinement method for camera centers and scene points, and leads to fast reconstruction of 3D scene points. We demonstrate the performance of our approach using several real-world and synthetic datasets and provide comparisons to other approaches.