Progressive Large-Scale Structure-from-Motion with Orthogonal MSTs

摘要

Pairwise image matching plays a vital role in Structure-from-Motion (SfM). Though the image-retrieval method accelerates the matching process, the number of neighbors is usually hard to determine. Insufficient feature matches could break the completeness of reconstructed scene, while redundant pairs may bring in many erroneous ones. In this paper, we propose a progressive SfM method to tackle the completeness, robustness and efficiency problems in a united framework, where two loops are contained. The outer loop is a feature matching loop, where the orthogonal MSTs (maximum spanning trees) of the image similarity graph is iteratively selected to perform the image matching. The inner loop is an incremental camera calibration loop, where the initial camera poses in each iteration are inherited from those calibrated in the last one. By progressively performing the image matching and calibration, we find both loops converge fast and a large number of redundant pairs are excluded. Experiments demonstrate the superior performance of our method in terms of both efficiency and robustness on various image datasets, and our method also has a large potential to tackle the ambiguity problems in SfM.