RANSAC Back to SOTA: A Two-stage Consensus Filtering for Real-time 3D Registration

TL;DR

This work targets fast and robust correspondence-based 3D registration under heavy outliers. It introduces a two-stage consensus filtering (TCF) that sequentially applies one-point RANSAC, two-point RANSAC, and three-point RANSAC, followed by scale-adaptive Cauchy IRLS to estimate the rigid transform. By reducing sampling dimensions and pruning outliers early, the approach achieves substantial speedups while maintaining high recall and accuracy, validated on ETH and KITTI with up to three orders of magnitude faster runtimes than previous SOTA methods. The method can complement and boost existing PCR pipelines, enabling real-time registration in challenging, real-world scenarios.

Abstract

Correspondence-based point cloud registration (PCR) plays a key role in robotics and computer vision. However, challenges like sensor noises, object occlusions, and descriptor limitations inevitably result in numerous outliers. RANSAC family is the most popular outlier removal solution. However, the requisite iterations escalate exponentially with the outlier ratio, rendering it far inferior to existing methods (SC2PCR [1], MAC [2], etc.) in terms of accuracy or speed. Thus, we propose a two-stage consensus filtering (TCF) that elevates RANSAC to state-of-the-art (SOTA) speed and accuracy. Firstly, one-point RANSAC obtains a consensus set based on length consistency. Subsequently, two-point RANSAC refines the set via angle consistency. Then, three-point RANSAC computes a coarse pose and removes outliers based on transformed correspondence's distances. Drawing on optimizations from one-point and two-point RANSAC, three-point RANSAC requires only a few iterations. Eventually, an iterative reweighted least squares (IRLS) is applied to yield the optimal pose. Experiments on the large-scale KITTI and ETH datasets demonstrate our method achieves up to three-orders-of-magnitude speedup compared to MAC while maintaining registration accuracy and recall. Our code is available at https://github.com/ShiPC-AI/TCF.

Figures (9)

• Figure 1: Average registration recall and runtime for ETH's tree sequence. The runtime denotes the average time for a single registration. Both our method and MAC achieve a 100% recall. Remarkably, ours accomplishes this in just 59 ms, leading to a three-orders-of-magnitude improvement over MAC's 331587 ms.
• Figure 2: Registration recall of RANSAC at different iterations. We create 200 pairs of point clouds containing 2% inliers to evaluate the registration recall.
• Figure 3: Overall framework and outlier removal illustration. (a): Our method cascades one-point, two-point, and three-point RANSAC, followed by scale-adaptive Cauchy IRLS, each involving an iterative process. The raw correspondence $\mathcal{C}$ is progressively refined into subsets $\mathcal{I}^\prime$, $\mathcal{I}^{\prime\prime}$, and $\mathcal{I}^{\prime\prime\prime}$, satisfying $\mathcal{I}^{\prime\prime\prime} \subseteq \mathcal{I}^{\prime\prime} \subseteq \mathcal{I}^\prime \subseteq \mathcal{C}$. The scale-adaptive Cauchy IRLS computes the final 6-DoF pose from $\mathcal{I}^{\prime\prime\prime}$. (b): The blue and black points represent the 3D correspondences. The green lines indicate inliers and red lines represent outliers. IR stands for the inlier ratio.
• Figure 4: One-point consensus generation. Same-colored points indicate a point correspondence. Dotted black lines show correct edge correspondences, while red indicates erroneous ones. An edge correspondence is correct if their length discrepancy $< 2\tau$. For $(p_k, q_k)$, applying length consistency identifies a maximal consensus comprising four inliers and one outlier marked in yellow.
• Figure 5: Angular discrepancy. The circled letter signifies a point, and the same color denotes a point correspondence. Black dotted lines represent correct edge correspondences, i.e., $|p_1p_2-q_1q_2| < 2\tau$ and $|p_3p_2-q_3q_2|< 2\tau$. The yellow outlier produces triangles with two similar edge lengths but different shapes due to angle discrepancies.