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

Dongxu Zhang; Jihua Zhu; Shiqi Li; Wenbiao Yan; Haoran Xu; Peilin Fan; Huimin Lu

cs.CVMar 13, 2026

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

Authors: Dongxu Zhang, Jihua Zhu, Shiqi Li, Wenbiao Yan, Haoran Xu, Peilin Fan, Huimin Lu

Abstract

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.

Figures & tables

Fig. 1: The Inlier Ratio (IR) is plotted on the x-axis for 3DMatch and on the y-axis for 3DLoMatch. IGASA stands out by consistently achieving the highest IR.

Fig. 2: The overview of the proposed IGASA framework. The pipeline operates in three stages. (a) , the HPA module processes the source $P$ and target $Q$ inputs to construct a hierarchical feature pyramid ( $F_{ordinary},F_{minor},F_{primary}$ ) with progressively expanding receptive fields. (b) , the HCLA module performs coarse matching by utilizing the SGIRA to fuse global semantics with local geometry, followed by the SAIGA for feature refinement. (c) , the IGAR module executes the fine registration phase, where an iterative optimization loop (repeated $N$ times) dynamically updates correspondence weights to suppress outliers and estimate the final transformation $\{R,t\}$ .

Fig. 3: Details of Gated Fusion Mechanism.

F_{p}^{n}

and

F_{q}^{n}

are preprocessed, passed through parallel convolutional layers, combined to form

F_{t}

, and then refined by a residual adjustment to produce

G_{\text{fusion}}

for weighted fusion.

Fig. 4: The result of IGASA on 3D(Lo)Match. We visualized some of the scenes in the dataset. The first row shows the initial arrangement of the two point clouds in their raw, unaligned positions.The second row displays the registration result predicted by our model.

Feature Matching Recall (%) $\uparrow$
	3DMatch					3DLoMatch
# Samples	5000	2500	1000	500	250	5000	2500	1000	500	250
FCGF [ 33 ]	97.4	97.3	97.0	96.7	96.6	76.6	75.4	74.2	71.7	67.3
Predator [ 18 ]	96.6	96.6	96.5	96.3	95.3	78.6	76.7	75.7	75.7	72.9
YOHO [ 63 ]	98.2	97.6	97.5	97.7	96.0	79.4	78.1	76.3	73.8	69.1
CoFiNet [ 19 ]	98.1	98.3	98.1	98.2	98.3	83.1	83.5	83.3	83.1	82.6

TABLE I: Evaluation results of our method on 3DMatch and 3DLoMatch datasets. Optimal results are highlighted in bold, with runner-up results indicated by an underline.

Model	RTE (cm)	RRE (°)	RR (%)
3DFeat-Net [ 67 ]	25.9	0.25	96.0
FCGF [ 33 ]	9.5	0.30	96.6
Predator [ 18 ]	6.8	0.27	99.8
CoFiNet [ 19 ]	8.2	0.41	99.8
HRegNet [ 68 ]	12	0.29	99.7
GeoTransformer [ 8 ]	6.8	0.24	99.8

TABLE II: Registration performance on KITTI dataset.

Fig. 5: Qualitative registration results on the KITTI odometry benchmark. We visualize four representative driving scenes characterized by large spatial extents and varying structural distinctiveness. The first row displays the initial unaligned input pairs (Source: Blue, Target: Yellow), illustrating significant initial pose discrepancies. The second row shows the registration results achieved by IGASA. The red rectangular boxes highlight specific regions with sharp structural boundaries, such as road curbs and intersection corners. Note that despite the limited overlap and noise typical of outdoor LiDAR scans, our method precisely aligns fine-grained structures such as trees, vehicles, and road boundaries.

Model	RTE (m)	RRE (°)	RR (%)
FGR [ 70 ]	0.71	1.01	32.2
DCP [ 71 ]	1.09	2.07	56.8
IDAM [ 72 ]	0.47	0.79	88.0
FMR [ 73 ]	0.60	1.61	92.1
DGR [ 74 ]	0.21	0.48	98.4
HRegNet [ 68 ]	0.18	0.45	99.9

TABLE III: Registration performance on nuScenes dataset.

Fig. 6: Qualitative registration results on the nuScenes dataset. These scenes exhibit extreme sparsity and non-uniform point density, which pose significant challenges for correspondence estimation. The first row shows the raw, unaligned point clouds. The second row presents the aligned results from IGASA. The red rectangular boxes highlight specific regions containing aligned sparse geometric cues, such as ground scan lines and structural edges. It can be observed that IGASA robustly handles the sparse geometry, effectively aligning the ground plane and sparse vertical structures ( $e.g.,$ traffic signs), demonstrating its efficacy in low-texture environments.

Module			3DMatch			3DLoMatch
HPA	HCLA	IGAR	FMR	IR	RR	FMR	IR	RR
$\checkmark$	-	-	92.7	80.2	91.3	72.8	49.7	70.8
-	$\checkmark$	-	94.2	83.5	91.9	75.9	55.1	72.7
-	-	$\checkmark$	93.0	81.7	91.5	69.2	52.6	71.5
-	$\checkmark$	$\checkmark$	97.2	85.2	93.4	74.3	57.6	74.2
$\checkmark$	-	$\checkmark$	95.6	81.9	92.8	71.0	51.3	72.3

TABLE IV: Ablation study of key modules in the IGASA framework on the 3DMatch and 3DLoMatch datasets. The results are reported in percentage (%). Best performance is highlighted in bold. The baseline denotes the network without HCLA and IGAR.

Module		3DMatch
SGIRA	SAIGA	FMR	IR	RR
-	-	95.6	81.9	92.8
$\checkmark$	-	96.2	83.3	93.2
-	$\checkmark$	96.7	84.2	93.5
$\checkmark$	$\checkmark$	98.2	87.9	94.6

TABLE V: Ablation experiments of the tow components in HCLA module on 3DMatch dataset. The results are reported in percentage (%). Best performance is highlighted in bold.

$\mathcal{L}_{\text{mat}}$	$\mathcal{L}_{\text{key}}$	$\mathcal{L}_{\text{den}}$	FMR	IR	RR
$\checkmark$	-	-	93.4	74.9	91.5
-	$\checkmark$	-	92.8	72.2	90.3
-	-	$\checkmark$	93.0	71.5	90.7
-	$\checkmark$	$\checkmark$	95.8	72.5	91.3
$\checkmark$	-	$\checkmark$	96.4	75.6	92.4
$\checkmark$	$\checkmark$	-	96.3	77.1	92.1

TABLE VI: Impact of different loss function combinations on registration performance.

\mathcal{L}_{\text{mat}}

\mathcal{L}_{\text{key}}

, and

\mathcal{L}_{\text{den}}

denote the matching, keypoint, and dense registration losses, respectively. The results are reported in percentage (%). Best performance is highlighted in bold.

Methods	Model time	Pose time	Total time
FCGF [ 33 ]	0.073	4.332	4.405
D3Feat [ 75 ]	0.035	4.472	4.506
SpinNet [ 76 ]	84.569	3.583	88.152
Predator [ 18 ]	0.050	7.014	7.064
CoFiNet [ 19 ]	0.167	2.493	2.660
GeoTransformer [ 8 ]	0.124	2.557	2.701

TABLE VII: Computational efficiency comparison on the 3DMatch benchmark. Times are measured in seconds (s). The total time represents the sum of feature extraction (Model time) and transformation estimation (Pose time).

Explore similar work

CMHANet: A Cross-Modal Hybrid Attention Network for Point Cloud Registration
Mar 13, 2026Dongxu Zhang, Yingsen Wang, Yiding Sun +3Point 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
GLASS: Geometry-aware Local Alignment and Structure Synchronization Network for 2D-3D Registration
Mar 27, 2026Zhixin Cheng, Jiacheng Deng, Xinjun Li +5Point Cloud Registration Cross-Modal Image Registration
School of Information Science and Technology, University of Science and Technology of China, Hefei 230027, China · Meituan Inc, Beijing 100000, China · Institute of Advanced Technology, University of Science and Technology of China, Hefei 230027, China
GeGS-PCR: Effective and Robust 3D Point Cloud Registration with Two-Stage Color-Enhanced Geometric-3DGS Fusion
Apr 20, 2026Jiayi Tian, Haiduo Huang, Tian Xia +2Point Cloud Registration
National Key Laboratory of Human-Machine Hybrid Augmented Intelligence, National Engineering Research Center of Visual Information and Applications, Institute of Artificial Intelligence and Robotics, Xi’an Jiaotong University, Xi’an, Shaanxi, China

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

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

Dongxu Zhang, Yingsen Wang, Yiding Sun, Haoran Xu+2

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

Point Cloud Registration

Open paper AskPDF arXiv

Robust point cloud registration is a fundamental task in 3D computer vision and geometric deep learning, essential for applications such as large-scale 3D reconstruction, augmented reality, and scene understanding. However, the performance of established learning-based methods often degrades in complex, real world scenarios characterized by incomplete data, sensor noise, and low overlap regions. To address these limitations, we propose CMHANet, a novel Cross-Modal Hybrid Attention Network. Our method integrates the fusion of rich contextual information from 2D images with the geometric detail of 3D point clouds, yielding a comprehensive and resilient feature representation. Furthermore, we introduce an innovative optimization function based on contrastive learning, which enforces geometric consistency and significantly improves the model's robustness to noise and partial observations. We evaluated CMHANet on the 3DMatch and the challenging 3DLoMatch datasets. \rev{Additionally, zero-shot evaluations on the TUM RGB-D SLAM dataset verify the model's generalization capability to unseen domains.} The experimental results demonstrate that our method achieves substantial improvements in both registration accuracy and overall robustness, outperforming current techniques. We also release our code in https://github.com/DongXu-Zhang/CMHANet.

Abstract

Explore similar work

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

GLASS: Geometry-aware Local Alignment and Structure Synchronization Network for 2D-3D Registration

GeGS-PCR: Effective and Robust 3D Point Cloud Registration with Two-Stage Color-Enhanced Geometric-3DGS Fusion