A Robust and Accurate Post-Validation Voting Scheme for Ranking 3D Correspondences

As a key step of the high-level 3D vision perception, extracting reliable inlier correspondences from the random correspondences with a high fraction of outliers is a challenging task. In this paper, we present a robust two-stage voting scheme to assess and rank the initial correspondences. Specifically, we first generate a geometric consistency point pair voting set by enforcing point pair feature constraints (PPFC) on the pairwise correspondences from an elected correspondence subset. Then, by combining the point pair voting set and the initial correspondence set, the compatible correspondence triples that satisfy PPFC are elaborately picked up to estimate the corresponding hypothesis poses, resulting in the pose voting set. Finally, by performing a kernel density estimation technique on the pose voting set, an optimal post-validation pose is estimated to compute the likelihood scores for the initial correspondence set. Extensive experiments on two challenging datasets demonstrate that the proposed voting scheme is robust against over 99% outlier ratio, outperforming the state-of-the-art methods in terms of inlier precision.