Hybrid Attention for Low-overlap Point Cloud Registration

doi:10.16182/j.issn1004731x.joss.25-1261

Abstract

Abstract:

To address the significant registration errors caused by large initial pose deviations in low overlap ratio point cloud, this paper presents a two-stage optimization method that integrates self-attention and cross-attention mechanisms. Orientation feature representation is enhanced using equivariant self-attention, semantic alignment is achieved through invariant cross-attention, and rotational equivariance is modeled with an octahedral group for coarse registration. Various attention variants are employed to aggregate features along the anchor dimension, and a local-to-global module is utilized for fine registration. On the 3DLoMatch dataset, the proposed method achieves a registration recall of 77.0% and an inlier ratio of 44.3%. Even with the introduction of random rotations, it maintains a registration recall of 72.0%, and under varying point cloud densities, the recall remains above 75.0%. Experimental results demonstrate that the method effectively improves registration accuracy and rotational robustness in low-overlap scenarios through equivariant feature extraction and multi-attention fusion.

Key words: point cloud registration, self-attention, cross-attention, low overlap ratio, deep learning

CLC Number:

TP391

Wang Huaijun, Yu Zonghua, Xu Chenxin, Li Junhuai, Li Chenzhi, Yin Junjie. Hybrid Attention for Low-overlap Point Cloud Registration[J]. Journal of System Simulation, 2026, 38(9): 2502-2517.

Figures/Tables 18

Table 1

Main notations

符号	定义与说明
P, Q	源点云与目标点云
N, M	原始源点云与目标点云的点数量
T	刚性变换矩阵
R, t	旋转矩阵与平移向量
$p i, q j$	点云P和Q中的三维坐标点
G	离散旋转群(本文采用正八面体群)
g	离散旋转群G中的一个特定旋转操作
A	正八面体离散旋转群的锚点数量
Nʹ, Mʹ	降采样后的点云点数量
C	特征通道维度
$X P, X Q$	源点云与目标点云的等变特征
$W Q, W K, W V$	注意力机制中的可训练权重矩阵
Rep(·)	点云旋转特征置换操作函数
$α$	注意力权重分数

Table 1

Fig. 1

Fig. 2

Fig. 3

Fig. 4

Fig. 5

Fig. 6

Fig. 7

Table 2

Table 3

Table 4

Table 5

Table 6

Fig. 8

Table 7

Table 8

Table 9

Table 10

References 41

[1]	鲁斌, 王明晗, 孙洋, 等. 全局信息感知与局部特征融合的高效三维目标检测[J]. 系统仿真学报, 2024, 36(11): 2616-2630.
	Lu Bin, Wang Minghan, Sun Yang, et al. Global-local Fusion for Efficient 3D Object Detection[J]. Journal of System Simulation, 2024, 36(11): 2616-2630.
[2]	王若萱, 吴建平, 徐辉. 自动驾驶汽车感知系统仿真的研究及应用综述[J]. 系统仿真学报, 2022, 34(12): 2507-2521.
	Wang Ruoxuan, Wu Jianping, Xu Hui. Overview of Research and Application on Autonomous Vehicle Oriented Perception System Simulation[J]. Journal of System Simulation, 2022, 34(12): 2507-2521.
[3]	Salas-Moreno R F, Newcombe R A, Strasdat H, et al. SLAM++: Simultaneous Localisation and Mapping at the Level of Objects[C]//2013 IEEE Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2013: 1352-1359.
[4]	李海英, 徐浩男, 郝俊芳. 面向三维场景的电力设备高斯溅射建模研究[J]. 系统仿真学报, 2026, 38(3): 595-607.
	Li Haiying, Xu Haonan, Hao Junfang. Research on Gaussian Splatting Modeling of Power Equipment in 3D Scenes[J]. Journal of System Simulation, 2026, 38(3): 595-607.
[5]	周伟, 苍慜楠, 程浩宗. 基于AR技术的文物数字化三维图像重建方法[J]. 图学学报, 2025, 46(2): 369-381.
	Zhou Wei, Cang Minnan, Cheng Haozong. Research on the Method of 3D Image Reconstruction for Cultural Relics Based on AR Technology[J]. Journal of Graphics, 2025, 46(2): 369-381.
[6]	张顺利, 徐艳芝, 周明全, 等. 基于自适应邻域匹配的点云配准方法[J]. 计算机学报, 2019, 42(9): 2114-2126.
	Zhang Shunli, Xu Yanzhi, Zhou Mingquan, et al. Registration of Point Clouds Based on Matching of General Adaptive Neighborhood[J]. Chinese Journal of Computers, 2019, 42(9): 2114-2126.
[7]	Thomas Hugues, Qi C R, Deschaud Jean-Emmanuel, et al. KPConv: Flexible and Deformable Convolution for Point Clouds[C]//2019 IEEE/CVF International Conference on Computer Vision (ICCV). Piscataway: IEEE, 2019: 6410-6419.
[8]	Guo Menghao, Cai Junxiong, Liu Zhengning, et al. PCT: Point Cloud Transformer[J]. Computational Visual Media, 2021, 7(2): 187-199.
[9]	杜沅昊, 耿秀丽, 徐诚智, 等. 基于改进灰狼算法和自适应分裂KD-tree的点云配准方法[J]. 系统仿真学报, 2025, 37(2): 424-435.
	Du Yuanhao, Geng Xiuli, Xu Chengzhi, et al. Point Cloud Registration Method Based on Improved Grey Wolf Algorithm and Adaptive Splitting KD-tree[J]. Journal of System Simulation, 2025, 37(2): 424-435.
[10]	Koide Kenji, Yokozuka Masashi, Oishi Shuji, et al. Voxelized GICP for Fast and Accurate 3D Point Cloud Registration[C]//2021 IEEE International Conference on Robotics and Automation (ICRA). Piscataway: IEEE, 2021: 11054-11059.
[11]	Wei Jiangxia, Cheng Lan, Hu Zhimin, et al. HPNet: High Precision Point Cloud Registration Using Feature Pyramid and Hybrid Position Encoding[J]. Computers & Graphics, 2024, 119: 103896.
[12]	Hu Han, Hou Yongkuo, Ding Yulin, et al. V2PNet: Voxel-to-point Feature Propagation and Fusion That Improves Feature Representation for Point Cloud Registration[J]. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2023, 16: 5077-5088.
[13]	Zhao Guiyu, Du Zewen, Guo Zhentao, et al. VRHCF: Cross-source Point Cloud Registration via Voxel Representation and Hierarchical Correspondence Filtering[C]//2024 IEEE International Conference on Multimedia and Expo (ICME). Piscataway: IEEE, 2024: 1-6.
[14]	Mengjin Lü, Yang Jie, Qi Zhiquan, et al. Rigid Pairwise 3D Point Cloud Registration: A Survey[J]. Pattern Recognition, 2024, 151: 110408.
[15]	Slimani Karim, Achard Catherine, Tamadazte Brahim. RoCNet++: Triangle-based Descriptor for Accurate and Robust Point Cloud Registration[J]. Pattern Recognition, 2024, 147: 110108.
[16]	Fischer Kai, Simon Martin, Ölsner Florian, et al. StickyPillars: Robust and Efficient Feature Matching on Point Clouds using Graph Neural Networks[C]//2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2021: 313-323.
[17]	Chen Suyi, Xu Hao, Li Haipeng, et al. PointRegGPT: Boosting 3D Point Cloud Registration Using Generative Point-cloud Pairs for Training[C]//Computer Vision — ECCV 2024. Cham: Springer Nature Switzerland, 2025: 272-289.
[18]	Poiesi Fabio, Boscaini Davide. Learning General and Distinctive 3D Local Deep Descriptors for Point Cloud Registration[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023, 45(3): 3979-3985.
[19]	Zhao Bao, Liu Qiang, Wang Zihan, et al. HA-TiNet: Learning a Distinctive and General 3D Local Descriptor for Point Cloud Registration[J]. IEEE Transactions on Visualization and Computer Graphics, 2025, 31(9): 5507-5518.
[20]	Choy C, Park Jaesik, Koltun V. Fully Convolutional Geometric Features[C]//2019 IEEE/CVF International Conference on Computer Vision (ICCV). Piscataway: IEEE, 2019: 8957-8965.
[21]	Wang Haiping, Liu Yuan, Dong Zhen, et al. You Only Hypothesize Once: Point Cloud Registration with Rotation-equivariant Descriptors[C]//Proceedings of the 30th ACM International Conference on Multimedia. New York: ACM, 2022: 1630-1641.
[22]	Yu Hao, Qin Zheng, Hou Ji, et al. Rotation-invariant Transformer for Point Cloud Matching[C]//2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2023: 5384-5393.
[23]	Dias Pais G, Ramalingam S, Venu Madhav Govindu, et al. 3DRegNet: A Deep Neural Network for 3D Point Registration[C]//2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2020: 7191-7201.
[24]	Lee Junha, Kim Seungwook, Cho Minsu, et al. Deep Hough Voting for Robust Global Registration[C]//2021 IEEE/CVF International Conference on Computer Vision (ICCV). Piscataway: IEEE, 2021: 15974-15983.
[25]	Yu Hao, Li Fu, Saleh Mahdi, et al. CoFiNet: Reliable Coarse-to-fine Correspondences for Robust Point Cloud Registration[C]//Proceedings of the 35th International Conference on Neural Information Processing Systems. Red Hook: Curran Associates Inc., 2021: 23872-23884.
[26]	Qin Zheng, Yu Hao, Wang Changjian, et al. GeoTransformer: Fast and Robust Point Cloud Registration with Geometric Transformer[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023, 45(8): 9806-9821.
[27]	Chen Hong, Yan Pei, Xiang Sihe, et al. Dynamic Cues-assisted Transformer for Robust Point Cloud Registration[C]//2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2024: 21698-21707.
[28]	Lin T Y, Zhu Minghan, Ghaffari M. Lie Neurons: Adjoint-equivariant Neural Networks for Semisimple Lie Algebras[C]//Proceedings of the 41st International Conference on Machine Learning. Chia Laguna Resort: PMLR, 2024: 30529-30545.
[29]	Deng Congyue, Litany O, Duan Yueqi, et al. Vector Neurons: A General Framework for SO(3)-equivariant Networks[C]//2021 IEEE/CVF International Conference on Computer Vision (ICCV). Piscataway: IEEE, 2021: 12180-12189.
[30]	Chen Haiwei, Liu Shichen, Chen Weikai, et al. Equivariant Point Network for 3D Point Cloud Analysis[C]//2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2021: 14509-14518.
[31]	Zhu Minghan, Ghaffari M, Clark W A, et al. E2PN: Efficient SE(3) — Equivariant Point Network[C]//2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2023: 1223-1232.
[32]	Luo Shitong, Li Jiahan, Guan Jiaqi, et al. Equivariant Point Cloud Analysis via Learning Orientations for Message Passing[C]//2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2022: 18910-18919.
[33]	Zeng A, Song Shuran, Nießner M, et al. 3DMatch: Learning Local Geometric Descriptors from RGB-D Reconstructions[C]//2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2017: 199-208.
[34]	Huang Shengyu, Zan Gojcic, Usvyatsov Mikhail, et al. PREDATOR: Registration of 3D Point Clouds with Low Overlap[C]//2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2021: 4265-4274.
[35]	Geiger Andreas, Lenz Philip, Urtasun R. Are We Ready for Autonomous Driving? The KITTI Vision Benchmark Suite[C]//2012 IEEE Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2012: 3354-3361.
[36]	Bai Xuyang, Luo Zixin, Zhou Lei, et al. PointDSC: Robust Point Cloud Registration Using Deep Spatial Consistency[C]//2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2021: 15854-15864.
[37]	Li Yang, Harada Tatsuya. Lepard: Learning Partial Point Cloud Matching in Rigid and Deformable Scenes[C]//2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2022: 5544-5554.
[38]	Yu Hao, Hou Ji, Qin Zheng, et al. Riga: Rotation-invariant and Globally-aware Descriptors for Point Cloud Registration[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024, 46(5): 3796-3812.
[39]	Bai Xuyang, Luo Zixin, Zhou Lei, et al. D3Feat: Joint Learning of Dense Detection and Description of 3D Local Features[C]//2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2020: 6358-6366.
[40]	Ao Sheng, Hu Qingyong, Yang Bo, et al. SpinNet: Learning a General Surface Descriptor for 3D Point Cloud Registration[C]//2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2021: 11748-11757.
[41]	Wang Haiping, Liu Yuan, Hu Qingyong, et al. RoReg: Pairwise Point Cloud Registration with Oriented Descriptors and Local Rotations[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023, 45(8): 10376-10393.

配置	规格
操作系统	Windows 10
GPU	NVIDIA GeForce RTX 4090 D
Python	3.8.20
CUDA	11.8
PyTorch	2.0.0

方法	随机旋转前			随机旋转后
方法	RR	IR	FMR	RR	IR	FMR
GeoTransformer^[26]	75.0	43.5	88.3	71.8	40.0	85.8
Lepard^[37]	65.4	28.4	83.1	49.0	24.4	79.5
RIGA^[38]	65.1	32.1	85.1	66.9	32.1	84.5
CoFiNet^[25]	67.5	24.4	83.1	62.5	21.5	78.6
FCGF^[20]	40.1	21.4	76.6	40.8	21.6	75.4
D3Feat^[39]	37.2	13.2	67.3	36.1	13.5	67.6
Predator^[34]	59.8	26.7	78.6	57.7	26.2	75.7
SpinNet^[40]	59.8	20.5	75.3	58.1	20.1	75.3
YOHO^[21]	65.2	25.9	79.4	65.9	26.4	79.2
RoReg^[41]	70.3	39.6	82.1	—	—	—
RoITr^[22]	74.8	54.3	89.6	—	—	—
本文	77.0	44.3	88.8	72.0	43.6	88.0

评估指标	方法	下采样点数量
评估指标	方法	5 000	2 500	1 000	500	250
RR	RoReg^[41]	70.3	70.2	68.3	67.6	64.9
	GeoTransformer^[26]	75.0	74.4	74.7	73.7	73.2
	CoFiNet^[25]	67.2	66.2	68.5	66.1	61.3
	Predator^[34]	60.2	61.1	62.7	61.3	57.4
	YOHO^[21]	65.2	65.5	63.2	56.5	48.0
	本文	77.0	77.0	76.2	75.6	75.0
IR	RoReg^[41]	39.6	39.9	33.6	32.0	34.5
	GeoTransformer^[26]	43.5	47.9	54.0	56.6	58.4
	CoFiNet^[25]	24.4	25.9	26.7	26.8	27.1
	Predator^[34]	26.7	28.0	28.3	25.7	25.7
	YOHO^[21]	25.9	23.3	22.6	18.2	15.0
	本文	44.3	50.2	54.8	56.7	57.9
FMR	RoReg^[41]	82.1	82.3	81.2	80.7	80.5
	GeoTransformer^[26]	88.3	87.2	87.7	87.8	87.3
	CoFiNet^[25]	83.1	83.1	82.9	82.7	82.7
	Predator^[34]	78.6	79.3	79.4	79.2	77.8
	YOHO^[21]	79.4	78.1	76.3	73.8	69.1
	本文	88.8	88.6	88.6	88.5	87.0

评估指标	方法	下采样点数量
评估指标	方法	5 000	2 500	1 000	500	250
RR	GeoTransformer^[26]	71.8	67.8	68.0	66.7	66.5
	YOHO^[21]	65.9	—	—	—	—
	本文	72.0	71.1	71.3	70.6	69.7
IR	GeoTransformer^[26]	40.0	44.1	50.7	53.6	55.5
	YOHO^[21]	26.4	—	—	—	—
	本文	43.6	49.5	55.1	57.5	59.0
FMR	GeoTransformer^[26]	85.8	86.1	86.0	86.0	85.9
	YOHO^[21]	79.2	—	—	—	—
	本文	88.0	88.4	88.5	88.1	87.2

方法	RRE/(°)	RTE/cm	RR/%
FCGF^[20]	0.30	9.5	96.6
D3Feat^[39]	0.30	7.2	99.8
SpinNet^[40]	0.47	9.9	99.1
Predator^[34]	0.27	6.8	99.8
CoFiNet^[25]	0.41	8.2	99.8
GeoTransformer^[26]	0.24	6.8	99.8
本文	0.22	6.7	99.8