Ranking 3D feature correspondences via consistency voting

This paper addresses the 3D correspondence ordering problem by proposing a consistency voting (CV) method. Given a set of plausible feature correspondences generated via geometric feature matching, our goal is to rank these correspondences accurately and robustly such that reasonable correspondences can be located according to their rankings. The proposed method employs a voting-based scheme, which is based on the consistency check of each initial correspondence with a predefined voting set, to calculate a voting score for each correspondence. We consider two consistency terms, i.e., rigidity and local reference frame (LRF) affinity, to ensure a correct judge when two correspondences are compatible. LRF is an intrinsic canonical cue widely used in many existing local geometric features. The repeatability of LRFs helps to remove the ambiguity that sometimes arises from the rigidity check. The proposed CV method is evaluated on three public benchmarks including both LiDAR and Kinect data. Comparison with the state-of-the-art confirms the effectiveness and robustness of the CV method with respect to noise, varying point densities, clutter, occlusion, and changes in data modality. Furthermore, the proposed CV method can rank thousands of correspondences in real time.