Abstract:Addressing the challenges posed by the substantial gap in point cloud data collected from diverse sensors, achieving robust cross-source point cloud registration becomes a formidable task. In response, we present a novel framework for point cloud registration with broad applicability, suitable for both homologous and cross-source registration scenarios. To tackle the issues arising from different densities and distributions in cross-source point cloud data, we introduce a feature representation based on spherical voxels. Furthermore, addressing the challenge of numerous outliers and mismatches in cross-source registration, we propose a hierarchical correspondence filtering approach. This method progressively filters out mismatches, yielding a set of high-quality correspondences. Our method exhibits versatile applicability and excels in both traditional homologous registration and challenging cross-source registration scenarios. Specifically, in homologous registration using the 3DMatch dataset, we achieve the highest registration recall of 95.1% and an inlier ratio of 87.8%. In cross-source point cloud registration, our method attains the best RR on the 3DCSR dataset, demonstrating a 9.3 percentage points improvement. The code is available at

What problem does this paper attempt to address?

This paper attempts to address the problem of robust and accurate cross-source point cloud registration between point cloud data from different sensor sources. Specifically, the paper mainly addresses the following two key challenges: 1. **Inconsistent Density and Distribution Differences**: Point cloud data from different sensors exhibit significant differences in density and distribution. For example, point clouds generated by LiDAR are usually sparse and have a linear structure, while point clouds generated by depth cameras are denser. These differences make it difficult to extract similar features through feature metric learning, thereby affecting the accuracy of registration. 2. **A Large Number of Outliers and Mismatches**: Point cloud data collected by different sensors contain a large number of outliers, which leads to a large number of mismatches during the feature matching process. These mismatches further increase the difficulty of cross-source registration. To address these issues, the paper proposes a new framework that includes the following key components: - **Feature Representation Based on Spherical Voxels**: By using spherical voxelization and multi-scale spherical normalization methods, density-invariant feature representation is achieved, avoiding feature loss caused by downsampling. - **Soft Correspondence Generation**: By calculating the distance matrix in the feature space and applying the softmax operation to generate soft correspondences, as many correct correspondences as possible are captured. - **Hierarchical Correspondence Filtering**: Through a multi-level correspondence filtering method, mismatches are gradually removed, and high-quality correspondences are finally obtained. Experimental results show that the proposed method performs excellently in cross-source point cloud registration tasks, especially outperforming existing methods on the 3DCSR dataset. Additionally, the method also achieves good results in same-source point cloud registration tasks, particularly achieving the highest registration recall rate and inlier ratio on the 3DMatch dataset.

VRHCF: Cross-Source Point Cloud Registration via Voxel Representation and Hierarchical Correspondence Filtering

Cross-PCR: A Robust Cross-Source Point Cloud Registration Framework

A Coarse-to-Fine Algorithm for Matching and Registration in 3D Cross-Source Point Clouds

Efficient Single Correspondence Voting for Point Cloud Registration

A Systematic Approach for Cross-source Point Cloud Registration by Preserving Macro and Micro Structures

Fast Registration for cross-source point clouds by using weak regional affinity and pixel-wise refinement

A Novel Method for Registration of MLS and Stereo Reconstructed Point Clouds

Robust point cloud registration using Hough voting-based correspondence outlier rejection

A Hierarchical Multiview Registration Framework of TLS Point Clouds Based on Loop Constraint

Color Point Cloud Registration Based on Supervoxel Correspondence

HybridFusion: LiDAR and Vision Cross-Source Point Cloud Fusion

Robust real-world point cloud registration by inlier detection

A Coarse-to-Fine Algorithm for Registration in 3D Street-View Cross-Source Point Clouds

VRNet: Learning the Rectified Virtual Corresponding Points for 3D Point Cloud Registration

Hierarchical Optimization of 3D Point Cloud Registration

Multi-instance Point Cloud Registration by Efficient Correspondence Clustering

Street view cross-sourced point cloud matching and registration

Practical, Fast and Robust Point Cloud Registration for 3D Scene Stitching and Object Localization

Point cloud registration and localization based on voxel plane features

Robust Point Cloud Registration Framework Based on Deep Graph Matching

An Automatic and Robust Point Cloud Registration Framework Based on View-Invariant Local Feature Descriptors and Transformation Consistency Verification