HybridPoint: Point Cloud Registration Based on Hybrid Point Sampling and Matching

Yiheng Li, Canhui Tang, Runzhao Yao, Aixue Ye, Feng Wen, Shaoyi Du

Introduction

Point cloud registration is a critical task in 3D computer vision, involving the alignment of multiple 3D scans into a unified coordinate system. Recent "patch-to-point" frameworks have gained popularity, typically partitioning point clouds into patches centered around sampled nodes, often referred to as superpoints. However, traditional grid-sampled superpoints often suffer from low repeatability and poor localization precision. In featureless regions or areas with low overlap, these superpoints fail to provide consistent anchors, leading to ambiguous correspondences and registration failure.

Methodology

The proposed HybridPoint framework addresses these limitations by introducing a hybrid sampling strategy and a dual-class matching module designed to improve both the repeatability and the coverage of sampled nodes.

Hybrid Point Sampler

The core innovation lies in the Hybrid Point Sampler, which categorizes points into two distinct types to balance precision and scene coverage:

  1. Salient Points: Extracted using Intrinsic Shape Signatures (ISS) and refined via Non-Maximum Suppression (NMS), these points represent high-curvature features like corners and edges. They offer high repeatability across different viewpoints, serving as reliable anchors.
  2. Non-Salient Points: To ensure global coverage, especially in planar or featureless regions where salient points are absent, the model incorporates uniformly distributed points.

Together, these form "hybrid points" that provide a more consistent partition of the 3D space compared to standard grid-based sampling.

Comparison of superpoint vs. hybrid point grouping. Hybrid points provide significantly better partition consistency between scans.

Dual-Classes Patch Matching

Recognizing that salient and non-salient points possess different geometric properties, HybridPoint employs a dual-branch matching strategy. The salient branch focuses on high-confidence geometric anchors that are inherently easier to match. The non-salient branch, which is more susceptible to noise due to the lack of distinct local geometry, utilizes a Spectral Matching (SM) algorithm. This algorithm enforces 3D spatial consistency among neighboring points, effectively filtering out outlier correspondences in featureless areas.

Architecture Overview

The framework utilizes a KPConv-FPN backbone for robust hierarchical feature extraction. A Geometric Transformer is integrated to learn rotation-invariant, context-aware descriptors, facilitating accurate matching even under significant pose variations. The final stage involves a repeatable point matching module and Local Geometric Registration (LGR) to produce the final alignment.

The HybridPoint architecture, detailing the feature extraction, hybrid sampling, and dual-class matching stages.

Experimental Results

HybridPoint was evaluated on the 3DMatch, 3DLoMatch, and KITTI Odometry benchmarks, achieving state-of-the-art (SOTA) performance across all metrics.

  • 3DMatch/3DLoMatch: On the 3DMatch dataset (high overlap), the model achieved a Registration Recall (RR) of 93.0%, surpassing previous SOTA methods like GeoTransformer (91.5%). In the more challenging 3DLoMatch scenario (low overlap), it reached 75.0% RR.
  • KITTI: The model demonstrated exceptional precision on large-scale outdoor LiDAR data, achieving 99.8% RR with significantly lower rotation and translation errors than competing models.

Qualitative analysis confirms that the hybrid sampling approach leads to more consistent patch partitions and a higher ratio of correct (inlier) correspondences compared to standard superpoint methods, particularly in environments dominated by planar surfaces.

Conclusion

HybridPoint demonstrates that the quality of initial point sampling is paramount for successful 3D registration. By combining repeatable salient anchors with comprehensive non-salient coverage and employing a specialized dual-class matching strategy, the framework sets a new standard for accuracy and robustness. Future work may explore the integration of semantic information to further refine patch partitioning and matching logic.

Summary updated

BibTeX
@inproceedings{li2023hybridpoint,
  author = {Yiheng Li and Canhui Tang and Runzhao Yao and Aixue Ye and Feng Wen and Shaoyi Du},
  title = {{HybridPoint: Point Cloud Registration Based on Hybrid Point Sampling and Matching}},
  booktitle = {ICME},
  year = {2023},
  doi = {10.1109/icme55011.2023.00346},
}

Loading related content...

Figure