Cross-view Domain Generalization via Geometric Consistency for LiDAR Semantic Segmentation

Jindong Zhao; Yuan Gao; Yang Xia; Sheng Nie; Jun Yue; Weiwei Sun; Shaobo Xia

cs.CVFeb 16, 2026

Cross-view Domain Generalization via Geometric Consistency for LiDAR Semantic Segmentation

Authors: Jindong Zhao, Yuan Gao, Yang Xia, Sheng Nie, Jun Yue, Weiwei Sun, Shaobo Xia

Organizations: School of Aeronautic Engineering, Changsha University of Science and Technology, Hunan, China, 410004. · Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing, China, 100094. · School of Automation, Central South University, Changsha 410083, China · University of British Columbia.

Abstract

Domain-generalized LiDAR semantic segmentation (LSS) seeks to train models on source-domain point clouds that generalize reliably to multiple unseen target domains, which is essential for real-world LiDAR applications. However, existing approaches assume similar acquisition views (e.g., vehicle-mounted) and struggle in cross-view scenarios, where observations differ substantially due to viewpoint-dependent structural incompleteness and non-uniform point density. Accordingly, we formulate cross-view domain generalization for LiDAR semantic segmentation and propose a novel framework, termed CVGC (Cross-View Geometric Consistency). Specifically, we introduce a cross-view geometric augmentation module that models viewpoint-induced variations in visibility and sampling density, generating multiple cross-view observations of the same scene. Subsequently, a geometric consistency module enforces consistent semantic and occupancy predictions across geometrically augmented point clouds of the same scene. Extensive experiments on six public LiDAR datasets establish the first systematic evaluation of cross-view domain generalization for LiDAR semantic segmentation, demonstrating that CVGC consistently outperforms state-of-the-art methods when generalizing from a single source domain to multiple target domains with heterogeneous acquisition viewpoints. The source code will be publicly available at https://github.com/KintomZi/CVGC-DG

Figures & tables

Figure 1: Challenges in cross-view domain generalization. (a) Conventional DG assumes similar sensing viewpoints across domains, whereas (b) Cross-view DG involves drastically different acquisition geometries, leading to severe density and visibility shifts. We address this challenge by generating view-dependent geometric variants and enforcing consistency across them.

Figure 2: Overview of the CVGC framework . The framework consists of two main modules: (i) Cross-view Geometric Augmentation (CGA), which generates a view-dependent variant $\mathbf{P}^{\alpha}$ from the source point cloud $\mathbf{P}^{s}$ ; and (ii) Geometric Consistency Regularization (GCR), which enforces structural and semantic consistency across variants. A shared backbone predicts point-wise semantics for both views, with semantic consistency and voxel occupancy supervision jointly regularizing the learned representations, using only source-domain data.

Figure 3: Illustration of the CGA. (a) Viewpoint-agnostic density resampling, which adjusts point cloud scene density via tangent plane-based densification and voxel-based sparsification. (b) Viewpoint-dependent visibility simulation, which constructs occluded scenes based on a spherical projection process.

Figure 4: Schematic of the occupancy perception head. Point-wise features are aggregated into voxel representations via voxelization and KNN-based interpolation, and processed by sparse convolutional layers to predict voxel occupancy, providing unified geometric supervision across different viewpoints.

Figure 5: Benchmark for cross-view domain generalization. The six datasets are partitioned into two groups following a one-source, two-target protocol. (a) First-Group Datasets: H3D (UAV) $\rightarrow$ Paris-Lille-3D (MLS) and ISPRS Vaihingen (ALS). (b) Second-Group Experiments: STPLS3D (synthetic UAV) $\rightarrow$ Toronto-3D (MLS) and DALES (ALS).

Figure 6: Qualitative results for Hessigheim-3D → Paris-Lille-3D. Results of different methods are visualized on the MLS target domain, where colors indicate semantic categories.

Source	Target	Method	Per-class IoU (%)					mIoU
Source	Target	Method	Ground	Building	Natural	Vehicle	U.Furniture	mIoU
Hessigheim-3D	Paris-Lille-3D	Source-only	91.98	83.83	82.55	50.65	21.30	66.06
		CoSMix [ 30 ]	95.29	87.05	80.04	64.71	31.93	71.80
		PolarMix [ 28 ]	94.20	82.75	62.13	29.90	6.97	55.19
		DGLSS [ 10 ]	92.33	84.89	81.12	42.58	20.23	64.23
		UniMix [ 11 ]	93.44	69.99	62.12	44.06	25.04	58.93

Table I: Quantitative comparison on Hessigheim-3D (Source) to Paris-Lille-3D and ISPRS Vaihingen (Targets) cross-view benchmarks.

Figure 7: Qualitative results for Hessigheim-3D → ISPRS Vaihingen. Results of different methods are visualized on the ALS target domain, where colors indicate semantic categories.

Figure 8: Qualitative results for STPLS3D → Toronto-3D. Results of different methods are visualized on the MLS target domain, where colors indicate semantic categories.

Figure 9: Qualitative results for STPLS3D → DALES. Results of different methods are visualized on the ALS target domain, where colors indicate semantic categories.

Source	Target	Method	Per-class IoU (%)						mIoU
Source	Target	Method	Ground	Building	Natural	Vehicle	Pole	Fence	mIoU
STPLS3D	Toronto-3D	Source-only	96.51	57.83	84.51	9.44	35.72	2.70	47.79
		CoSMix [ 30 ]	89.73	17.96	32.43	6.80	15.34	0.93	27.20
		PolarMix [ 28 ]	89.45	45.08	62.58	13.09	28.34	3.68	40.37
		DGLSS [ 10 ]	96.13	60.15	81.63	12.18	35.90	3.41	48.23
		UniMix [ 11 ]	95.85	41.87	85.19	32.05	31.74	3.08	48.30

Table II: Quantitative comparison on STPLS3D (Source) to Toronto-3D and DALES (Targets) cross-view benchmarks.

Source	Target	Module		mIoU
Source	Target	CGA	GCR	mIoU
Hessigheim-3D	Paris-Lille-3D	✗	✗	66.06
		✓	✗	67.04
		✓	✓	72.87
	ISPRS Vaihingen	✗	✗	31.20
		✓	✗	53.17

Table III: Ablation study on Hessigheim-3D as source, with Paris-Lille-3D and ISPRS Vaihingen as targets.

Source	Target	Module		mIoU
Source	Target	CGA	GCR	mIoU
STPLS3D	Toronto-3D	✗	✗	47.79
		✓	✗	50.42
		✓	✓	51.28
	DALES	✗	✗	58.57
		✓	✗	61.91

Table IV: Ablation study on STPLS3D as source, with Toronto-3D and DALES as targets.

Domain Generalization

Domain generalization develops models that perform reliably in new environments without training on target-domain data.

See all 375 papers

CVSD-Reg: Cross-Modal Visual Semantic Prior Distillation for Robust LiDAR Registration
Aug 20, 2026Eunsoo Im, Junghun Suh, Gyeonggwan Lee +1Point Cloud Registration Point Cloud Learning
LiFi: LiDAR Generation from Multi-View Images via Geometric and Semantic Collaborative Guidance
Oct 14, 2026Sizhuo Zhou, Xiaosong Jia, Fanrui Zhang +10Point Cloud Generation Diffusion Model Guidance
University of Science and Technology of China · Shanghai Innovation Institute · Fudan University +1
OmniLiDAR: A Unified Diffusion Framework for Multi-Domain 3D LiDAR Generation
May 13, 2026Youquan Liu, Weidong Yang, Ao Liang +9Synthetic Data Generation for Robotics 3D Scene Generation
College of Computer Science and Artificial Intelligence, Fudan University, Shanghai, China · School of Computing, Department of Computer Science, National University of Singapore, Singapore · College of Computer Science and Technology, Nanjing University of Aeronautics and Astronautics, Nanjing, China +5

Aug 20, 2026J/K move · Enter open · S save

CVSD-Reg: Cross-Modal Visual Semantic Prior Distillation for Robust LiDAR Registration

Eunsoo Im, Junghun Suh, Gyeonggwan Lee, Seunghwan Hong

Point Cloud Registration Point Cloud Learning LiDAR Localization Cross-Modal Knowledge Distillation

Open paper AskPDF arXiv

Learning-based global point cloud registration has achieved remarkable progress, yet its reliance on geometric representations makes existing methods sensitive to variations in point density, scan pattern, viewpoint, and sensor characteristics. We propose CVSD-Reg, a robust global LiDAR registration framework that distills visual semantic priors from a vision foundation model into LiDAR representations. In Stage 1, a Point Transformer V3 student learns from a frozen DINOv2 teacher through contrastive distillation and spherical-manifold alignment, which preserves the hyperspherical geometry of the teacher embedding space. Self-supervised InfoNCE consistency and soft

\mathrm{SE}(3)

invariance further encourage viewpoint-robust descriptors. In Stage 2, the distilled representation is adapted to registration through correspondence learning, density-aware point-dropout augmentation, and end-to-end pose optimization. With a single checkpoint, CVSD-Reg generalizes to both single-sensor and zero-shot cross-sensor scenarios without sensor-specific adaptation and remains entirely camera-free at inference. On KITTI, nuScenes, and HeLiPR, CVSD-Reg achieves strict success rate ([email protected],m/

1^\circ

) of 97.7

\%

, 99.0

\%

, and 99.3

\%

, respectively, including 97.3

\%

on sparse 16-beam Velodyne scans. It outperforms state-of-the-art geometric registration methods by up to 44.0 percentage points without requiring camera inputs or post-hoc ICP refinement.

Abstract

Domain Generalization

Explore similar work

CVSD-Reg: Cross-Modal Visual Semantic Prior Distillation for Robust LiDAR Registration

LiFi: LiDAR Generation from Multi-View Images via Geometric and Semantic Collaborative Guidance

OmniLiDAR: A Unified Diffusion Framework for Multi-Domain 3D LiDAR Generation