Progressive Correspondence Regenerator for Robust 3D Registration

TL;DR

Regor tackles robust 3D registration under extreme outliers by shifting from pruning to progressive regeneration of correspondences. It integrates prior-guided local grouping, generalized mutual matching, and a center-aware three-point consistency refinement within local regions, followed by a global refinement stage to produce dense, high-quality matches and accurate 6DoF pose estimation. The approach is supported by theoretical guarantees (Theorem 1 and Theorem 2) and strong empirical results on 3DMatch, 3DLoMatch, and KITTI, including extreme outlier scenarios where traditional methods fail, and demonstrates robustness across both traditional and learning-based descriptors. By regenerating inliers rather than discarding them, Regor achieves significantly higher inlier counts and registration reliability, offering a practical path toward robust 3D registration in challenging real-world conditions.

Abstract

Obtaining enough high-quality correspondences is crucial for robust registration. Existing correspondence refinement methods mostly follow the paradigm of outlier removal, which either fails to correctly identify the accurate correspondences under extreme outlier ratios, or select too few correct correspondences to support robust registration. To address this challenge, we propose a novel approach named Regor, which is a progressive correspondence regenerator that generates higher-quality matches whist sufficiently robust for numerous outliers. In each iteration, we first apply prior-guided local grouping and generalized mutual matching to generate the local region correspondences. A powerful center-aware three-point consistency is then presented to achieve local correspondence correction, instead of removal. Further, we employ global correspondence refinement to obtain accurate correspondences from a global perspective. Through progressive iterations, this process yields a large number of high-quality correspondences. Extensive experiments on both indoor and outdoor datasets demonstrate that the proposed Regor significantly outperforms existing outlier removal techniques. More critically, our approach obtain 10 times more correct correspondences than outlier removal methods. As a result, our method is able to achieve robust registration even with weak features. The code will be released.

Figures (12)

• Figure 1: Difference between our method and existing outlier removal techniques. Outlier removal methods perform top-down pruning, producing an optimized set of correspondences that is a subset of the initial correspondences. When the initial inliers are few, this approach can obtain only a limited number of inliers, which is insufficient for robust pose estimation. In contrast, our method takes a bottom-up approach, using a regeneration strategy to generate more inliers, thereby effectively addressing this issue.
• Figure 2: Overall framework of our method. We extract features from the original point cloud, obtaining features $\mathbf{{F} }^{\mathcal{P}}$ and $\mathbf{{F} }^{\mathcal{Q}}$ as input for our method. Subsequently, a progressive process is applied to iteratively regenerate more accurate and denser correspondences $\mathcal{{G}}^{t}$. At each iteration, the output correspondences $\mathcal{{G}}^{t-1}$ from the previous iteration serve as input. Firstly, prior-guided local grouping is employed to sample seed corresponding points and form local correspondence regions $\mathbf{P}^{t}_i$ and $\mathbf{Q}^{t}_i$. Then, for each pair of local correspondence regions, generalized mutual matching is performed to get new correspondences. Next, these correspondences are refined locally and globally using our center-aware three-point consistency, followed by a merging operation $\oplus$ of local correspondences $\mathcal{{G}}^{t}_{i}$ using a hash table. Finally, using these refined correspondences, we achieve robust and accurate transformation estimation $\mathbf T\{\mathbf R,\mathbf t\}$ only with SVD.
• Figure 3: Illustration of GMM. The orange, blue, green, and red lines are the MNN correspondences, NN correspondences, correct correspondences, and wrong correspondences, respectively.
• Figure 4: Illustration of the local correspondence correction. Yellow and blue areas represent a pair of local point clouds. The purple line indicates the correct correspondence after correction.
• Figure 5: Correspondences before and after processing.