| Abstract | Recovering a 3D scene from unordered photo collections is a long-studied topic in computer vision. Existing reconstruction pipelines, both incremental and global, have already achieved remarkable results. This paper addresses the problem of fusing multiple existing partial 3D reconstructions, in particular finding the overlapping regions and transformations (7 DOF) between partial reconstructions. Unlike the previous methods which have to take the entire epipolar geometry (EG) graph as the input and reconstruct the scene, we propose an approach that reuses the existing reconstructed 3D models as input and merges them by utilizing all the internal information to avoid repeated work. This approach is divided into two steps. The first is to find overlapping areas between partial reconstructions based on Fisher similarity lists. Then, based on those overlaps, pairwise rotation between partial reconstructions is estimated by solving an `1 approximation optimization problem. After global rotation estimation, translation and scale between each pair of partial reconstructions are computed simultaneously in a global manner. In order to find the optimal transformation path, the maximal spanning tree (MST) is constructed in the second stage. Our approach is evaluated on diverse challenging public datasets and compared to state-of-the-art Structure from Motion (SfM) methods. Experiments show that our merging approach achieves high computational efficiency while preserving similar reconstruction accuracy and robustness. In addition, our method has superior extensibility which can add partial 3D reconstructions gradually to extend an existing 3D scene. |
|---|
