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

TL;DR

This work presents a novel framework for point cloud registration with broad applicability, suitable for both homologous and cross-source registration scenarios, and introduces a feature representation based on spherical voxels to tackle the issues arising from different densities and distributions in cross-source point cloud data.

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 https://github.com/GuiyuZhao/VRHCF.

Figures (6)

• Figure 1: Challenges in cross-source point cloud registration.
• Figure 2: Given cross-source point clouds $\mathbf{P}$ and $\mathbf{Q}$, we first obtain local patches $\mathbf{P}_i$ and $\mathbf{Q}_j$ by FPS qi2017pointnet and KNN search, and then complete the representation of initial features $\mathbf{\widetilde{F} }^{\mathcal{P}_i}$ and $\mathbf{\widetilde{F} }^{\mathcal{Q}_j}$ by spherical voxelization and multi-scale sphere normalization. Then, the features are refined by the 3DCCN backbone to obtain the features $\mathbf{\widehat{F} }^{\mathcal{P}_i}$ and $\mathbf{\widehat{F} }^{\mathcal{Q}_j}$. Soft correspondence generation is applied to loosely generate the initial correspondences $\mathcal{G}$, which are further filtered through hierarchical correspondence filtering to produce the final correspondence $\mathcal{G}^L$. The transformation $\mathbf{T}$ is ultimately determined using SVD or RANSAC fischler1981random.
• Figure 3: The superiority of our normalization method where the green circle denotes the area where the point distribution of the two patches is the same, and the red indicates a difference. (For brevity, the split in the radius direction is plotted.)
• Figure 4: Qualitative results. The correct correspondences are denoted by the green lines, while the red lines indicate errors.
• Figure 5: qualitative results on Kinect-SFM.