R-SLPR: Region-based Small-to-Large Point-cloud Registration

Abstract

Point-cloud (PC) registration is fundamental to three-dimensional (3D) perception in robotic systems. However, classic registration algorithms falter when aligning a source PC containing limited, incomplete, or ambiguous geometric cues against a reference. This challenge of registering a small, partial PC to a significantly larger global reference is pervasive in real-world deployment yet remains insufficiently addressed by existing learning-based approaches, which typically assume comparable scales and significant overlap.

To bridge this gap, we propose the Region-based Small-to-Large Point-cloud Registration framework (R-SLPR), a novel three-stage architecture that fundamentally reformulates the scale-mismatched registration problem into a sequence of region proposal, regional matching, and iterative refinement. Unlike conventional methods that fail to localize specific regions, R-SLPR explicitly identifies candidate regions prior to estimating rigid transformations, ensuring robust alignment even under severe scale mismatch. The framework introduces a Fibonacci Grid Segmentation method coupled with a contrastive learning objective to effectively generate and match local geometric patches. Building on this, a novel Cascade Anchor Selection and Refinement algorithm iteratively aligns the source with the target region to maximize precision. Extensive evaluation on ModelNet40 demonstrates that R-SLPR establishes a new state-of-the-art accuracy standard, outperforming prior approaches and significantly reducing position and rotation Mean Absolute Error (MAE) to 0.009 and 1.104, respectively.

Main Contributions

Region-based SLPR

We formulate small-to-large point cloud registration as a region proposal, matching, and refinement problem. By localizing candidate regions before rigid transformation estimation, R-SLPR reduces the ambiguity caused by severe scale mismatch and limited overlap.

Contrastive Patch Matching

We introduce a patch-level contrastive learning strategy to distinguish corresponding regions from irrelevant ones. This improves the discriminability of local geometric features under incomplete and ambiguous observations.

Cascade Region Refinement

We develop a Cascade Anchor Selection and Refinement module that iteratively evaluates and updates regional proposals. This allows R-SLPR to progressively converge toward the correct region and improve registration accuracy.

Method Overview

Overview of the proposed R-SLPR framework.

R-SLPR is designed for small-to-large point-cloud registration, where a small and incomplete source point cloud must be aligned with a much larger reference point cloud. Instead of directly matching the source to the entire reference, R-SLPR first identifies candidate local regions in the reference point cloud and then estimates the rigid transformation within the selected region.

The framework consists of three main steps: region proposal generation, proposal matching, and iterative refinement. First, the reference point cloud is segmented into multiple local proposals using Fibonacci Grid Segmentation. Then, a learning-based registration model evaluates the relationship between the source point cloud and each proposal, while contrastive learning improves the discriminability of regional features. Finally, Cascade Anchor Selection and Refinement updates the proposal anchors and gradually guides the search toward the correct matching region.

Key Components

Fibonacci Grid Segmentation

Fibonacci Grid Segmentation generates uniformly distributed anchor directions on the unit sphere and uses them to divide the large reference point cloud into regional proposals. Compared with direct global matching, this provides a more focused search space and allows the model to locate the region that is most likely to correspond to the small source point cloud.

Contrastive Learning

The contrastive learning module trains the feature extractor with source, reference, and irrelevant point-cloud patches. It pulls corresponding source-reference regions closer in the feature space and pushes unrelated regions apart, making the learned features more suitable for distinguishing candidate regions under severe scale mismatch.

Cascade Anchor Selection and Refinement

During inference, each regional proposal is evaluated by its predicted alignment quality. The best proposal is selected according to the average alignment distance, and its anchor direction is used to refine the proposal set in the next iteration. This cascade process progressively narrows the search area and improves registration robustness.

Experimental Results

R-SLPR achieves state-of-the-art performance on the ModelNet40 benchmark for small-to-large point cloud registration. As shown in the comparison tables, R-SLPR consistently outperforms representative handcrafted and learning-based registration methods across clean, unseen-category, noisy, and noisy-unseen settings. Under the clean setting with a sampling rate of 0.2, R-SLPR achieves an RMSE-r of 2.632°, an MAE-r of 1.104°, an RMSE-t of 0.020, and an MAE-t of 0.009, obtaining the best overall performance among all compared methods.

Compared with previous methods, R-SLPR demonstrates clear advantages in both rotation and translation estimation. In particular, the improvement in translation accuracy is especially significant, as conventional registration methods often suffer from centroid inconsistency when directly aligning a small source point cloud with a much larger reference point cloud. By first identifying the corresponding local region in the reference point cloud and then estimating the rigid transformation, R-SLPR effectively reduces the impact of scale mismatch and enables more accurate small-to-large registration.

The robustness of R-SLPR is further verified under challenging noisy and unseen-category conditions. While several baseline methods show substantial degradation or fail to converge under noise, R-SLPR maintains stable registration accuracy across all evaluated settings. These results demonstrate that the proposed region-based formulation provides a reliable and generalizable solution for small-to-large point cloud registration.

Clean Setting on ModelNet40
Method	Sampling Rate 0.2				Sampling Rate 0.3
Method	RMSE-r (°) ↓	MAE-r (°) ↓	RMSE-t ↓	MAE-t ↓	RMSE-r (°) ↓	MAE-r (°) ↓	RMSE-t ↓	MAE-t ↓
RPMNet	4.102	2.223	0.065	0.020	3.841	2.011	0.039	0.026
MFGNet	29.158	18.802	0.450	0.304	13.272	10.366	0.180	0.135
RORNet	18.635	15.314	0.322	0.240	16.469	12.391	0.340	0.271
MCLNet	3.556	1.132	0.224	0.156	3.429	1.186	0.168	0.092
R-SLPR (ours)	2.632	1.104	0.020	0.009	3.304	1.268	0.027	0.012

Unseen-Category Setting on ModelNet40
Method	Sampling Rate 0.2				Sampling Rate 0.3
Method	RMSE-r (°) ↓	MAE-r (°) ↓	RMSE-t ↓	MAE-t ↓	RMSE-r (°) ↓	MAE-r (°) ↓	RMSE-t ↓	MAE-t ↓
RPMNet	4.368	2.371	0.071	0.042	4.129	2.413	0.056	0.035
MFGNet	30.556	20.222	0.237	0.173	11.521	7.072	0.090	0.060
RORNet	19.913	14.260	0.366	0.204	16.395	12.856	0.410	0.306
MCLNet	3.638	1.328	0.150	0.098	3.521	1.200	0.140	0.101
R-SLPR (ours)	2.833	1.234	0.026	0.012	2.977	1.183	0.023	0.010

Noisy Setting on ModelNet40
Method	Sampling Rate 0.2				Sampling Rate 0.3
Method	RMSE-r (°) ↓	MAE-r (°) ↓	RMSE-t ↓	MAE-t ↓	RMSE-r (°) ↓	MAE-r (°) ↓	RMSE-t ↓	MAE-t ↓
RPMNet	4.562	2.740	0.072	0.025	3.985	2.219	0.047	0.019
MFGNet	33.340	21.842	0.404	0.320	14.395	9.567	0.168	0.124
RORNet	10.254	7.352	0.224	0.156	7.856	5.421	0.208	0.149
MCLNet	-	-	-	-	-	-	-	-
R-SLPR (ours)	2.613	1.142	0.027	0.013	2.974	1.229	0.021	0.010

Noisy & Unseen-Category Setting on ModelNet40
Method	Sampling Rate 0.2				Sampling Rate 0.3
Method	RMSE-r (°) ↓	MAE-r (°) ↓	RMSE-t ↓	MAE-t ↓	RMSE-r (°) ↓	MAE-r (°) ↓	RMSE-t ↓	MAE-t ↓
RPMNet	4.714	2.828	0.041	0.022	4.114	2.263	0.033	0.019
MFGNet	80.267	61.713	0.424	0.348	14.952	10.697	0.127	0.084
RORNet	11.026	7.864	0.271	0.197	7.882	5.663	0.325	0.188
MCLNet	-	-	-	-	-	-	-	-
R-SLPR (ours)	2.851	1.101	0.031	0.014	2.673	1.174	0.028	0.011

Data

The links to the datasets used in this paper are provided below.

ModelNet40 Data ISOPR Data

BibTeX

@inproceedings{wan2026rslpr,
      title     = {R-SLPR: Region-based Small-to-Large Point-cloud Registration with Contrastive Learning},
      author    = {Wan, Yusen and Chen, Zeyuan and Zou, Qianshi and Chen, Xu},
      booktitle = {Underreview},
      year      = {2026},
    }

R-SLPR: Region-based Small-to-Large Point-cloud Registration with Contrastive Learning

R-SLPR addresses small-to-large point-cloud registration by identifying candidate regions before estimating rigid transformations.