Introduction
Point cloud registration—the process of aligning multiple 3D scans into a unified coordinate system—is a fundamental challenge in robotics, autonomous driving, and computer vision. While traditional methods relied on handcrafted local features that offered high precision but struggled with noise, modern deep learning approaches have shifted toward high-level features that capture global context. However, this shift often comes at the cost of local geometric accuracy.
HD2Reg addresses this trade-off by introducing a hierarchical framework that explicitly decouples local geometric details from global semantic features. By employing a novel keypoint ranking strategy and a coarse-to-fine matching scheme, HD2Reg achieves state-of-the-art accuracy and robustness, particularly in challenging low-overlap scenarios.
The Problem: The Blur of Implicit Hierarchy
Most contemporary deep learning models for registration fuse local and global features implicitly, often through skip connections or multi-scale architectures. This "blurring" of levels results in representations that are neither distinctive enough globally to handle low-overlap cases nor precise enough locally to ensure fine-grained alignment. HD2Reg identifies this gap and proposes a solution through the explicit extraction and supervision of dual-level descriptors and detectors.
Methodology
The HD2Reg framework is built upon three core technical innovations designed to preserve the strengths of both local and global information.
1. Hierarchical Contrastive Learning
Rather than training a single monolithic descriptor, HD2Reg outputs two distinct representations:
- Low-level Descriptors: Focused on immediate local neighborhoods to capture fine geometric detail.
- High-level Descriptors: Focused on global distinctiveness to capture the broader context of the scene.
During training, these are supervised via hierarchical contrastive learning. Low-level features are pushed to be unique relative to their immediate neighbors, while high-level features are trained to be unique across the entire point cloud, ensuring that each level of the hierarchy serves its specific purpose.
2. Dual Keypoint Ranking
Standard registration models typically use binary labels to determine if a keypoint is "matchable." HD2Reg replaces this coarse supervision with a 4-level ranking system:
- Robust & Accurate
- Robust but Inaccurate
- Accurate but Unrobust
- Unrobust & Inaccurate
This 2-bit ranking system provides a much finer supervision signal, allowing the model to prioritize keypoints that are both repeatable across scans and geometrically precise.
3. Global-to-Local Matching
The registration process follows a coarse-to-fine pipeline. First, high-level features are utilized to establish coarse correspondences across the entire scene. Subsequently, the model performs local matching within small "cells" surrounding these initial points, using high-precision low-level features to "snap" the alignment into its final, accurate position.
Experimental Results
The authors evaluated HD2Reg against top-tier baselines, including Predator, FCGF, and D3Feat, across three standard benchmarks: 3DMatch (indoor), 3DLoMatch (low-overlap indoor), and KITTI Odometry (outdoor LiDAR).
Robustness and Precision
HD2Reg demonstrated significant improvements in challenging scenarios. On the 3DLoMatch dataset (10-30% overlap), it achieved a Registration Recall of approximately 64%, outperforming Predator’s 60%. In indoor scenes, the model consistently reduced Relative Rotation Error (RRE) and Relative Translation Error (RTE) across various sampling densities.
Outdoor Performance
On the KITTI dataset, HD2Reg achieved a 6.3 cm translation error, the lowest among compared keypoint-based methods, proving its efficacy in large-scale autonomous driving contexts.
Efficiency and Inlier Ratio
The "Global-to-Local" matching scheme significantly boosted the Inlier Ratio—the percentage of correct matches—reaching 69.6% on 3DMatch, compared to 54.1% for Predator. Furthermore, the method is remarkably robust to downsampling; even when using only 250 keypoints, HD2Reg maintains high registration recall, making it a strong candidate for real-time applications.
Conclusion and Future Work
HD2Reg demonstrates that explicit hierarchical separation is key to balancing robustness and precision in 3D registration. By moving away from binary keypoint labels to a nuanced ranking system, the model identifies more reliable correspondences even in low-overlap environments.
While the results are impressive, the authors note that the current pipeline requires a two-stage training process, which increases complexity. Additionally, maintaining dual branches for descriptors and detectors may increase the memory footprint. Future research may focus on streamlining this training process or extending the ranking strategy to other forms of 3D data beyond point clouds.
Summary updated