SANDRO: a Robust Solver with a Splitting Strategy for Point Cloud Registration

TL;DR

SANDRO addresses robust point cloud registration in challenging scenarios with high outlier rates and symmetries by marrying IRLS with a Geman-McClure robust loss under a Graduated Non-convexity framework. A novel splitting strategy partitions data into sub-clouds, mitigating initial bias and improving convergence without requiring initialization. Empirical results show up to 60% improvements on synthetic data and around 20% on Redwood compared with state-of-the-art methods, with online-friendly runtimes and options for fewer splits when speed is critical. The approach maintains robustness without discarding outliers, enabling reliable pose estimation in navigation and reconstruction tasks.

Abstract

Point cloud registration is a critical problem in computer vision and robotics, especially in the field of navigation. Current methods often fail when faced with high outlier rates or take a long time to converge to a suitable solution. In this work, we introduce a novel algorithm for point cloud registration called SANDRO (Splitting strategy for point cloud Alignment using Non-convex anD Robust Optimization), which combines an Iteratively Reweighted Least Squares (IRLS) framework with a robust loss function with graduated non-convexity. This approach is further enhanced by a splitting strategy designed to handle high outlier rates and skewed distributions of outliers. SANDRO is capable of addressing important limitations of existing methods, as in challenging scenarios where the presence of high outlier rates and point cloud symmetries significantly hinder convergence. SANDRO achieves superior performance in terms of success rate when compared to the state-of-the-art methods, demonstrating a 20% improvement from the current state of the art when tested on the Redwood real dataset and 60% improvement when tested on synthetic data.

Figures (7)

• Figure 1: Example of a symmetric relationship with initial bias. Point cloud 1 (left), is contained within point cloud 2 (right). The two chairs in point cloud 2 cause current registration methods to fail due to the initial outlier distribution. The matching chairs are shown with green boxes.
• Figure 2: Comparison of FPFH mutual matches between point clouds (left) and registered point clouds using SANDRO with the splitting strategy (right). Red and blue parts correspond to two different point clouds to be registered. In this example, the presence of two similar chairs in the point clouds causes the outlier distribution to be highly skewed, preventing convergence with traditional methods. The splitting strategy aims to break the initial bias in distribution by performing registration on independent subsets. The combined point cloud on the right shows the final registration of SANDRO with 4 splits.
• Figure 3: GNC applied to the Geman-McClure loss for different values of the decay parameter $\alpha$.
• Figure 4: Comparison of Rotation and Translation Errors, displayed only for the point cloud pairs which achieved a successful registration in the Redwood dataset.
• Figure 5: Aligned point clouds from the synthetic dataset with 50% outlier rate, representing a person standing. The points in red correspond to the source cloud, and the points in blue are the target cloud with outliers.