Abstract

This paper proposes a hierarchical framework that resamples 3D reconstructed points to reduce computation cost on time and memory for very large-scale Structure from Motion. The goal is to maintain ac- curacy and stability similar for different resample rates. We consider this problem in a level-of-detail perspective, from a very large scale global and sparse bundle adjustment to a very detailed and local dense optimization. The dense matching are resampled by exploring the redundancy using local invariant properties, while 3D points are resampled by exploring the redundancy using their covariance and their distribution in both 3D and image space. Detailed experiments on our resample framework are provided. We also demonstrate the proposed framework on large-scale ex- amples. The results show that the proposed resample scheme can produce a 3D reconstruction with the stability similar to quasi dense methods, while the problem size is as neat as sparse methods.