DualReg: Dual-Space Filtering and Reinforcement for Rigid Registration

Abstract

Noisy, partially overlapping data and the need for real-time processing pose major challenges for rigid registration. Considering that feature-based matching can handle large transformation differences but suffers from limited accuracy, while local geometry-based matching can achieve fine-grained local alignment but relies heavily on a good initial transformation, we propose a novel dual-space paradigm to fully leverage the strengths of both approaches. First, we introduce an efficient filtering mechanism consisting of a computationally lightweight one-point RANSAC algorithm and a subsequent refinement module to eliminate unreliable feature-based correspondences. Subsequently, we treat the filtered correspondences as anchor points, extract geometric proxies, and formulate an effective objective function with a tailored solver to estimate the transformation. Experiments verify our method's effectiveness, as demonstrated by a 32x CPU-time speedup over MAC on KITTI with comparable accuracy. Project page: https://ustc3dv.github.io/DualReg/.

Figures (9)

• Figure 1: Registration Recall and Average Runtime (per-registration) on the 3DMatch dataset with FPFH features. Optimal methods require high recall and low runtime. Our method achieves the highest recall with lower time cost.
• Figure 2: Pipeline of the proposed method. We propose DualReg, a dual-space paradigm for robust rigid registration. We first propose an efficient filtering algorithm for feature-based correspondences including 1-point RANSAC-based fast filtering and 3-point RANSAC-based refinement. Then, according to the filtered feature-based correspondences, we construct geometric proxies and propose a dual-space optimization framework to jointly estimate the rigid transformation. We mark the inliers and outliers with green and red lines, respectively, and mark the ground truth inlier rate (IR) at each stage.
• Figure 3: Correspondences meet symmetry.
• Figure 4: Iterative process of one-point RANSAC. Each circle in the figure represents a correspondence, labeled in the form of "$\mathbf{c}_j,s_{\mathbf{c}_j}^{(k)}$", where "$\mathbf{c}_j$" denotes the $j$-th correspondence and "$s_{\mathbf{c}_j}^{(k)}$" represents the confidence score of $\mathbf{c}_j$ at $k$-iteration. During the iteration, a correspondence is randomly selected each time to construct its consensus set: if there is a line between two correspondences, it indicates that they satisfy both length consistency and normal consistency; correspondences covered with the same color belong to the same consensus set.
• Figure 5: Qualitative comparison on 3DMatch. The yellow and blue point clouds represent the source and target point clouds, respectively. The first two rows correspond to FPFH-based examples with a ground truth inlier ratio of 1.55%, while the latter two rows demonstrate FCGF-based cases exhibiting a ground truth inlier ratio of 6.39%. Rotation error (RE) and translation error (TE) are marked, with red and green representing failed and successful registration, respectively.