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

Pengcheng Shi, Shaocheng Yan, Yilin Xiao, Xinyi Liu, Yongjun Zhang, Jiayuan Li

Abstract

3D point cloud registration (PCR) remains a significant challenge in high-outlier environments. While RANSAC is the industry standard, its computational cost grows exponentially as outlier ratios increase. This paper introduces the Two-stage Consensus Filtering (TCF) framework, which revitalizes RANSAC by employing cascaded one-point and two-point geometric filters. TCF achieves state-of-the-art accuracy with a 1000x speedup over existing robust methods, enabling real-time performance even when outliers exceed 95%.

Introduction

Aligning 3D scans into a unified coordinate system is foundational for autonomous navigation and robotic mapping. However, sensor noise and repetitive patterns often generate "outliers"—incorrect point correspondences. Traditional RANSAC handles outliers by randomly sampling subsets of matches, but as outlier ratios climb, the number of iterations required for a successful alignment becomes prohibitive. Modern alternatives like MAC or TEASER offer robustness but often at the cost of significant latency, sometimes taking minutes to process a single pair of scans.

Methodology: The TCF Framework

The TCF framework addresses the sampling bottleneck by decomposing the registration process into stages of increasing geometric complexity. Instead of immediately attempting a full 6-DOF pose estimation with three points, TCF prunes the correspondence set using simpler, computationally "cheap" constraints.

The TCF framework (top) and the progressive outlier removal process (bottom), showing the Inlier Ratio (IR) increasing from 5.3% to 100%.
  1. One-point RANSAC: This stage uses a single match and a "length consistency" constraint. It verifies if the distance between two points in the source cloud matches the distance between their counterparts in the target cloud, quickly identifying a large consensus set.
  2. Two-point RANSAC: Building on the first stage, this step applies "angle consistency." It ensures that triangles formed by matches are congruent across both clouds, filtering out outliers that survived the initial length check.
  3. Three-point RANSAC: With the vast majority of outliers removed, standard RANSAC requires only a few iterations to find a high-quality coarse pose.
  4. Scale-adaptive Cauchy IRLS: A final Iteratively Reweighted Least Squares optimization fine-tunes the 6-DOF pose for maximum precision.

Experimental Results

The TCF framework was evaluated on the ETH and KITTI datasets. The ETH "trees" sequence is particularly notorious for its complex geometry and extremely low inlier ratios (often near 1%).

Recall vs. Runtime on the ETH tree sequence. TCF (red star) achieves 100% recall significantly faster than competitors like MAC and TEASER.

The results demonstrate a massive performance leap. On the ETH trees sequence, TCF achieved 100% recall in just 59 ms. In contrast, the previous state-of-the-art, MAC, required over 331 seconds—a speedup of over three orders of magnitude. TCF maintained near-perfect recall even when the initial inlier ratio was as low as 1.1%, proving its extreme robustness across both urban and natural environments.

Conclusion

TCF demonstrates that classic geometric principles, when structured into efficient filtering stages, can outperform modern graph-based and deep-learning approaches in both speed and reliability. By lowering the sampling dimension in the early stages of RANSAC, TCF provides a "plug-and-play" solution for real-time, robust 3D registration in the wild.

Summary updated

BibTeX
@article{shi2024ransac,
  author = {Pengcheng Shi and Shaocheng Yan and Yilin Xiao and Xinyi Liu and Yongjun Zhang and Jiayuan Li},
  title = {{RANSAC Back to SOTA: A Two-stage Consensus Filtering for Real-time 3D Registration}},
  journal = {IEEE Robotics Autom. Lett.},
  year = {2024},
  url = {https://openreview.net/forum?id=iFj4Rp8Sng},
}
Figure