R3PM-Net: Real-time, Robust, Real-world Point Matching Network

Yasaman Kashefbahrami; Erkut Akdag; Panagiotis Meletis; Evgeniya Balmashnova; Dip Goswami; Egor Bondarau

cs.CVApr 6, 2026

R3PM-Net: Real-time, Robust, Real-world Point Matching Network

Authors: Yasaman Kashefbahrami, Erkut Akdag, Panagiotis Meletis, Evgeniya Balmashnova, Dip Goswami, Egor Bondarau

Organizations: Department of Electrical Engineering, Eindhoven University of Technology · Sioux Technologies

Abstract

Accurate Point Cloud Registration (PCR) is an important task in 3D data processing, involving the estimation of a rigid transformation between two point clouds. While deep-learning methods have addressed key limitations of traditional non-learning approaches, such as sensitivity to noise, outliers, occlusion, and initialization, they are developed and evaluated on clean, dense, synthetic datasets (limiting their generalizability to real-world industrial scenarios). This paper introduces R3PM-Net, a lightweight, global-aware, object-level point matching network designed to bridge this gap by prioritizing both generalizability and real-time efficiency. To support this transition, two datasets, Sioux-Cranfield and Sioux-Scans, are proposed. They provide an evaluation ground for registering imperfect photogrammetric and event-camera scans to digital CAD models, and have been made publicly available. Extensive experiments demonstrate that R3PM-Net achieves competitive accuracy with unmatched speed. On ModelNet40, it reaches a perfect fitness score of $1$ and inlier RMSE of $0.029$ cm in only $0.007$ s, approximately 7 times faster than the state-of-the-art method RegTR. This performance carries over to the Sioux-Cranfield dataset, maintaining a fitness of $1$ and inlier RMSE of $0.030$ cm with similarly low latency. Furthermore, on the highly challenging Sioux-Scans dataset, R3PM-Net successfully resolves edge cases in under 50 ms. These results confirm that R3PM-Net offers a robust, high-speed solution for critical industrial applications, where precision and real-time performance are indispensable. The code and datasets are available at https://github.com/YasiiKB/R3PM-Net.

Figures & tables

Figure 1: Qualitative results of R3PM-Net. As a real-time, feature-based deep learning method for point cloud registration, R3PM-Net improves generalizability and inference speed by expanding the network’s field-of-view.

Figure 2: Overview of the R3PM-Net Architecture. Built upon [ 34 ] , R3PM-Net employs an iterative Siamese framework for robust point cloud registration. The architecture includes four primary stages: (1) global-aware feature extraction employing shared MLPs to learn geometric similarities across a full receptive field; (2) correspondence estimation via a match matrix $M$ computed from feature distances; (3) outlier rejection driven by a parameter prediction module (4) transformation estimation using a differentiable SVD module to solve for the optimal rigid alignment $(R^{*},t^{*})$ . The estimated transformation is applied back to the source point cloud for iterative refinement.

Figure 3: Coarse-to-Fine Registration approach to handle noisy, sparse scans. Initially, $X$ and $Y$ are pre-processed via downsampling, normalization and alignment. Large misalignments are then resolved through coarse global registration performed by R3PM-Net, followed by a final high-precision local refinement with GICP.

Method	RRE [ ∘ ] $\downarrow$	RTE [cm] $\downarrow$	CD [cm] $\downarrow$	Fitness $\uparrow$	Inlier RMSE [cm] $\downarrow$	Time [s] $\downarrow$
RPMNet [ 34 ]	30.898 $\pm$ 0.322	0.002 $\pm$ 0.000	0.153 $\pm$ 0.001	0.998 $\pm$ 0.000	0.094 $\pm$ 0.001	0.021 $\pm$ 0.000
Predator [ 14 ]	7.262 $\pm$ 0.555	0.028 $\pm$ 0.002	0.045 $\pm$ 0.002	1.000 $\pm$ 0.000	0.026 $\pm$ 0.001	0.071 $\pm$ 0.001
GeoTransformer [ 18 ]	50.357 $\pm$ 1.238	0.215 $\pm$ 0.003	0.255 $\pm$ 0.005	0.921 $\pm$ 0.001	0.101 $\pm$ 0.002	0.065 $\pm$ 0.000
RegTR [ 35 ]	1.712 $\pm$ 0.061	0.007 $\pm$ 0.000	0.017 $\pm$ 0.000	1.000 $\pm$ 0.000	0.009 $\pm$ 0.000	0.045 $\pm$ 0.000
LoGDesc [ 22 ]	42.762 $\pm$ 1.254	0.158 $\pm$ 0.003	0.183 $\pm$ 0.004	0.978 $\pm$ 0.001	0.097 $\pm$ 0.002	0.075 $\pm$ 0.000
R3PM-Net (ours)	5.198 $\pm$ 0.067	0.010 $\pm$ 0.000	0.052 $\pm$ 0.000	1.000 $\pm$ 0.000	0.029 $\pm$ 0.000	0.007 $\pm$ 0.000

Table 1: Quantitative comparison on ModelNet40. Performance is reported in terms of RRE, RTE, Chamfer distance (CD), fitness (↑ higher is better), inlier RMSE, and inference time (↓ lower is better). R3PM-Net achieves the fastest inference while maintaining competitive accuracy. Best and second-best results are highlighted.

Method	RRE [ ∘ ] $\downarrow$	RTE [cm] $\downarrow$	CD [cm] $\downarrow$	Fitness $\uparrow$	Inlier RMSE [cm] $\downarrow$	Time [s] $\downarrow$
RPMNet [ 34 ]	32.217 $\pm$ 0.844	0.002 $\pm$ 0.000	0.160 $\pm$ 0.004	0.997 $\pm$ 0.002	0.098 $\pm$ 0.001	0.021 $\pm$ 0.000
Predator [ 14 ]	16.448 $\pm$ 1.548	0.044 $\pm$ 0.003	0.072 $\pm$ 0.002	1.000 $\pm$ 0.000	0.042 $\pm$ 0.001	0.071 $\pm$ 0.001
GeoTrans. [ 18 ]	45.582 $\pm$ 0.549	0.183 $\pm$ 0.005	0.297 $\pm$ 0.007	0.906 $\pm$ 0.004	0.111 $\pm$ 0.002	0.065 $\pm$ 0.000
RegTR [ 35 ]	1.311 $\pm$ 0.032	0.004 $\pm$ 0.000	0.023 $\pm$ 0.000	1.000 $\pm$ 0.000	0.012 $\pm$ 0.000	0.045 $\pm$ 0.000
LoGDesc [ 22 ]	121.224 $\pm$ 10.396	0.773 $\pm$ 0.060	0.692 $\pm$ 0.022	0.718 $\pm$ 0.010	0.224 $\pm$ 0.002	0.075 $\pm$ 0.000
R3PM-Net (ours)	5.451 $\pm$ 0.287	0.006 $\pm$ 0.001	0.054 $\pm$ 0.002	1.000 $\pm$ 0.000	0.030 $\pm$ 0.001	0.006 $\pm$ 0.000

Table 2: Quantitative comparison on Sioux-Cranfield dataset. R3PM-Net offers the best trade-off between real-time execution and competitive accuracy; it remains highly competitive, exceeding the performance of recent methods on most metrics while operating at an inference latency over

6.5\times

faster than RegTR. Best and Second-best results are highlighted.

Method	Lime			Cube			House			Teeth			Success	Time
Method	CD	Fit.	RMSE	CD	Fit.	RMSE	CD	Fit.	RMSE	CD	Fit.	RMSE	Rate	(s)
RPMNet [ 34 ]	0.270	1.000	0.047	0.290	1.000	0.023	-	-	-	-	-	-	28.6%	0.042
Predator [ 14 ]	0.270	1.000	0.048	0.289	1.000	0.023	-	-	-	-	-	-	28.6%	0.038
GeoTrans. [ 18 ]	0.260	1.000	0.041	0.295	1.000	0.024	-	-	-	-	-	-	28.6%	0.042
RegTR [ 35 ]	0.270	1.000	0.047	0.292	1.000	0.023	-	-	-	-	-	-	28.6%	0.038
LoGDesc [ 22 ]	-	-	-	0.292	1.000	0.024	0.222	1.000	0.052	-	-	-	28.6%	0.043

Table 3: Registration performance for successful cases of Sioux-Scans dataset (visually verified). ’-’ indicates that the method did not solve the object case. Success Rate is the ratio of successful trials to the seven test objects. Runtime reflects the average time (in seconds) to attempt all seven cases. Please refer to the Supplementary material (section A.3 ) for the full table including all failed cases.

Figure 4: Qualitative registration results of R3PM-Net on real-world event-camera data. It successfully aligns the “teeth” and “cube” models in under $50$ ms.

Model	Total Params. [M] $\downarrow$	Throughput [fps]
RPMNet	0.91	48
Predator	22.57	14
GeoTrans.	5.21	15
RegTR	11.49	22
LoGDesc	4.71	13
R3PM-Net (ours)	0.96	167

Table 4: Model complexity and throughput comparison. Throughput is the number of point cloud pairs registered per second. Despite comparable performance, R3PM-Net reduces total parameters by over

90\%

compared to RegTR.

Input Rep.	Norm.	Rad.	RRE $\downarrow$	RTE $\downarrow$	CD $\downarrow$	Fit. $\uparrow$	Time
PC Feat.1	✓	Fixed	31.86	0.00	0.16	1.00	0.021
PC Feat.2	–	Fixed	9.35	0.01	0.07	1.00	0.006
PC Feat.3	✓	Flex	13.65	0.01	0.08	0.99	0.006
PC Feat.4	–	Flex	12.90	0.01	0.08	0.99	0.006
Direct PC R3PM-Net	–	–	2.01	0.00	0.02	1.00	0.006

Table 5: Ablation study on input features. Hand-crafted local features are compared to R3PM-Net’s direct point cloud (PC) approach on the Sioux-Cranfield dataset. Bold indicates the best results.

Dataset	Method	RRE $\downarrow$	RTE $\downarrow$	CD $\downarrow$	Fit. $\uparrow$	Time
Model-Net40	R3PM-Net	5.198	0.010	0.052	0.029	0.007
	R3PM-Net (FT)	1.963	0.003	0.025	0.014	0.007
Sioux-Cranfield	R3PM-Net	5.451	0.006	0.054	0.030	0.006
	R3PM-Net (FT)	2.297	0.002	0.033	0.018	0.006

Table 6: Ablation study of cross-domain fine-tuning (FT). Fine-tuning on a subset of the proposed Sioux-Cranfield dataset improves performance across both dataset benchmarks. Bold indicates best results.

Method	FT Subset	Success Cases	Rate (%)
R3PM-Net	—	teeth, cube	28.6
R3PM-Net (FT)*	teeth, lime, cube, lego	teeth, lime, house	42.9
R3PM-Net (FT)	rd-peg, sep, shoe	teeth, lime, cube	42.9
	Plate, eleph., house	cube	14.3
	Plate, eleph., house, rd-peg	teeth, cube	28.6
	rd-peg, sep, shoe, lego	teeth	14.3

Table 7: Ablation study on fine-tuning subsets of Sioux-Cranfield dataset evaluated on Sioux-Scans.* denotes the model studied in the effectiveness of fine-tuning experiment shown in Table 6 .

Figure 5: 3D Quality Control Setup developed by Sioux Technologies [ 24 ] . This setup leverages event-camera technology to provide fast, high-fidelity 3D scanning suitable for in-line real-time quality assurance for specialized products. It aims to deliver optimized process control in applications including 3D metal printing, integrated electronics in molded parts, and dental prosthetics.

Category	Source Type	Qty
Sioux (Reconstructed)	Photogram.	4
Sioux (Synthetic)	CAD Models	3
Cranfield [ 6 ]	Pristine	6
Total	—	13

Table 8: Composition of the Sioux-Cranfield Dataset. Sioux (Reconstructed) and Sioux (Synthetic) are also used in Sioux-Scans dataset to produce Target point clouds.

Figure 6: CAD models of the Sioux-Cranfield dataset. The first six belong to the Cranfield Assembly benchmark [ 6 ] and the rest are contributions of this paper (Sioux dataset).

Figure 7: Sioux-Scans point cloud data overview. Target (blue) and Source (yellow) point clouds for seven distinct objects.

Metric	Type	Pros	Cons	GT-Free
RRE	Geodesic dist.	Scale-invariant	Less intuitive, not always objective	✗
RTE	Euclidean dist.	Scale-invariant	Less intuitive, not always objective	✗
Chamfer Dist.	Euclidean dist.	Suitable for symmetric shapes	Scale-variant	✓
Fitness	Overlap ratio	Intuitive	Scale-variant	✓
Inlier RMSE	Euclidean dist.	Practical	Scale-variant, can be misleading with poor alignment	✓
Time	Seconds	Important for real-time use	Does not reflect accuracy	✓

Table 9: Summary of Evaluation Metrics: Types, strengths, limitations, and ground-truth (GT) requirements. Scale-invariant metrics do not depend on the point cloud size. GT-free metrics are suitable for real-world scenarios whereas GT-based metrics are used in simulations.

Method	Data	Teeth	Lime	Cube	Lego	Eleph.	House	Shoe	SR (%)	Time
RPMNet [ 34 ]	CD	.205	.270	.290	.218	.173	.264	.191	28.6	0.042s
	Fit.	1.00	1.00	1.00	1.00	1.00	1.00	.968
	RMSE	.059	.047	.023	.065	.064	.132	.106
	Status	failed	success	success	failed	failed	failed	failed
Predator [ 14 ]	CD	.185	.270	.289	.210	.158	.284	.092	28.6	0.038s
	Fit.	1.00	1.00	1.00	1.00	1.00	.992	1.00

Table 10: Complete performance comparison of baseline methods on real-life industrial data. Results are averaged over seven independent runs across seven test objects. A successful registration is defined as achieving accurate alignment, verified visually, in at least four out of seven trials.

Point Cloud Registration

Point cloud registration aligns 3D point sets by estimating the transformations and correspondences that make their geometry consistent.

See all 110 papers

IGASA: Integrated Geometry-Aware and Skip-Attention Modules for Enhanced Point Cloud Registration
Mar 13, 2026Dongxu Zhang, Jihua Zhu, Shiqi Li +4Point Cloud Registration Point Cloud Learning
CMHANet: A Cross-Modal Hybrid Attention Network for Point Cloud Registration
Mar 13, 2026Dongxu Zhang, Yingsen Wang, Yiding Sun +33D Shape Matching Point Cloud Registration
School of Software, Xi’an Jiaotong University, No. 28 Xianning West Road, China · CEPRI, China Electric Power Research Institute No. 15 Xiaoying East Road, Qinghe, Haidian District, Beijing, China
PC2Model: ISPRS benchmark on 3D point cloud to model registration
Apr 21, 2026Mehdi Maboudi, Said Harb, Jackson Ferrao +3Sim-to-Real Transfer Point Cloud Registration
Institute of Geodesy and Photogrammetry, Technische Universität Braunschweig, Germany · Institute of Geodesy · Department of Infrastructure Engineering, University of Melbourne, Australia. +1

Mar 13, 2026J/K move · Enter open · S save

IGASA: Integrated Geometry-Aware and Skip-Attention Modules for Enhanced Point Cloud Registration

Dongxu Zhang, Jihua Zhu, Shiqi Li, Wenbiao Yan+3

Point Cloud Registration Point Cloud Learning Attention Mechanisms

Open paper AskPDF arXiv

Point cloud registration (PCR) is a fundamental task in 3D vision and provides essential support for applications such as autonomous driving, robotics, and environmental modeling. Despite its widespread use, existing methods often fail when facing real-world challenges like heavy noise, significant occlusions, and large-scale transformations. These limitations frequently result in compromised registration accuracy and insufficient robustness in complex environments. In this paper, we propose IGASA as a novel registration framework constructed upon a Hierarchical Pyramid Architecture (HPA) designed for robust multi-scale feature extraction and fusion. The framework integrates two pivotal components consisting of the Hierarchical Cross-Layer Attention (HCLA) module and the Iterative Geometry-Aware Refinement (IGAR) module. The HCLA module utilizes skip attention mechanisms to align multi-resolution features and enhance local geometric consistency. Simultaneously, the IGAR module is designed for the fine matching phase by leveraging reliable correspondences established during coarse matching. This synergistic integration within the architecture allows IGASA to adapt effectively to diverse point cloud structures and intricate transformations. We evaluate the performance of IGASA on four widely recognized benchmark datasets including 3D(Lo)Match, KITTI, and nuScenes. Our extensive experiments consistently demonstrate that IGASA significantly surpasses state-of-the-art methods and achieves notable improvements in registration accuracy. This work provides a robust foundation for advancing point cloud registration techniques while offering valuable insights for practical 3D vision applications. The code for IGASA is available in https://github.com/DongXu-Zhang/IGASA.

Abstract

Point Cloud Registration

Explore similar work

IGASA: Integrated Geometry-Aware and Skip-Attention Modules for Enhanced Point Cloud Registration

CMHANet: A Cross-Modal Hybrid Attention Network for Point Cloud Registration

PC2Model: ISPRS benchmark on 3D point cloud to model registration