Training-Free Global Geometric Association for 4D LiDAR Panoptic Segmentation

Gyeongrok Oh; Youngdong Jang; Jonghyun Choi; Suk-Ju Kang; Guang Lin; Sangpil Kim

cs.CVDec 22, 2025

Training-Free Global Geometric Association for 4D LiDAR Panoptic Segmentation

Authors: Gyeongrok Oh, Youngdong Jang, Jonghyun Choi, Suk-Ju Kang, Guang Lin, Sangpil Kim

Organizations: Korea University · Hyundai Motor Company · Sogang University · Purdue University

Abstract

Dominant paradigms for 4D LiDAR panoptic segmentation are usually required to train deep neural networks with large superimposed point clouds or design dedicated modules for instance association. However, these approaches perform redundant point processing and consequently become computationally expensive, yet still overlook the rich geometric priors inherently provided by raw point clouds. To this end, we introduce \textsc{Geo-4D}, a simple yet effective training-free framework that unifies spatial and temporal reasoning, enabling holistic LiDAR perception over long time horizons. Specifically, we propose a global geometric association strategy that establishes consistent instance correspondences by estimating an optimal transformation between instance-level point sets. To mitigate instability caused by structural inconsistencies in point cloud observations, we propose a global geometry-aware soft matching mechanism that enforces spatially coherent point-wise correspondences grounded in the spatial distribution of instance point sets. Furthermore, our carefully designed pipeline, which considers three instance types-static, dynamic, and missing-offers computational efficiency and occlusion-aware matching. Our extensive experiments across both SemanticKITTI and nuScenes demonstrate that our method consistently outperforms state-of-the-art approaches, even without additional training or extra point cloud inputs.

Figures & tables

Strategy	Frame-wise Perception	No Temporal Training	Off-the-shelf 3D Perception Model
IoU-based	✗	✗	✗
Query Prop.	✓	✗	✗
Detect & Track	✓	✗	✓
Geo-4D (Ours)	✓	✓	✓

Table 1 : Conceptual comparison of 4D LiDAR panoptic segmentation paradigms.

Figure 1 : (a) Nearest neighbor aligns each source point with its closest counterpart, focusing only on local point proximity. (b) In contrast, our Global-Geometry-aware Soft Matching computes correspondences via an optimal transport plan $\hat{\mathcal{Q}}$ , enabling instance-aware matching that respects the global geometry of each instance point set.

		Validation						Test		Efficiency
Method	#Sc.	LSTQ $\uparrow$	$S_{\rm assoc}\uparrow$	MOTSA $\uparrow$	sMOTSA $\uparrow$	PTQ $\uparrow$	sPTQ $\uparrow$	LSTQ $\uparrow$	$S_{\rm assoc}\uparrow$	Mem. $\downarrow$	Inf. Time $\downarrow$
4D-StOP ( 4 )	2	66.4	71.8	31.0	27.1	53.4	53.7	62.9	67.3	15.5	3.67
Eq-4D-StOP ( 5 )	2	68.9	74.8	43.6	39.6	56.4	56.7	65.4	69.7	11.0	2.89
Mask4Former ( 6 )	2	72.0	76.3	47.3	42.1	48.3	48.5	68.4	67.3	10.4	3.51
Mask4D ( 12 )	1	71.4	75.4	52.1	48.1	58.4	58.5	64.3	66.4	11.9	1.12
Mask4Former + CA-Net ( 14 )	1	70.4	73.4	48.4	43.1	48.9	49.1	66.5	62.8	5.71	1.56

Table 2 : Quantitative results on SemanticKITTI. We compare Geo-4D with prior methods in terms of the standard 4D panoptic metric (LSTQ), complementary tracking metrics, and computational efficiency. Memory usage is reported in GB, and inference time is reported in seconds per scan. Best and second best results are shown in bold and underlined , respectively.

Figure 2 : Qualitative comparison on SemanticKITTI . We visualize the association results over five consecutive scans for both the baselines and our Geo-4D . Different colors represent different instances. The dotted boxes zoom in for clear comparison. For clarity, we use circled numbers (①–⑤) to represent the frame indices.

Method	#Scan	LSTQ $\uparrow$	$S_{\rm assoc}\uparrow$
Centroid + Hungarian	1	55.4	45.4
3D Box IoU + Hungarian	1	63.6	59.9
Kalman + 3D Box IoU + Hungarian	1	64.6	61.9
Geo-4D (Ours)	1	72.7	78.3

Table 3 : Association analysis on SemanticKITTI. (a) Comparison with classic association baselines. (b) Controlled comparison under identical frame-wise predictions.

Method	#Sc.	LSTQ $\uparrow$	$S_{\rm assoc}\uparrow$	$S_{\rm cls}\uparrow$
4D-StOP ( 4 )	4	60.5	62.5	58.6
Eq-4D-StOP ( 5 )	2	58.9	62.3	55.7
Mask4Former ( 6 )	2	67.6	63.0	72.5
Mask4D ( 12 )	1	58.5	54.1	63.1
Mask4Former	1	–	–	71.0
+ CA-Net ( 14 )	1	63.4	56.6	71.0

Table 4 : Quantitative results on panoptic nuScenes ( 28 ) .

Figure 3 : Visualization of point-wise correspondences. (a) Nearest-neighbor matching. (b) Our Global Geometry-aware Soft Matching. $\blacksquare$ and ${\color[rgb]{0,0,1}\bullet}$ represent the target and source points, respectively.

Figure 4 : (a) Comparison with representative baselines across different association strategy categories on rigid and non-rigid object classes. (b) Class-wise performance with respect to object class types and the average number of points per instance.

	Components			Metric
Exp.	Hist. Init.	GGSM	Memory	LSTQ $\uparrow$	$S_{\rm assoc}\uparrow$
1	–	–	–	70.1	73.4
2	✓	–	–	71.3	75.4
3	–	✓	–	71.4	75.6
4	✓	✓	–	72.0	76.8
5	✓	✓	✓	72.7	78.3

Table 5 : (a) Effect of each module in Geo-4D . (b) Effect of handling different instance types.

Figure 5 : We illustrate the performance variation w.r.t. the hyperparameters of Geo-4D . The dashed line denotes the average performance of three different association methods: Mask4Former ( 6 ) , Mask4D ( 12 ) , and CA-Net ( 14 ) .

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Rank	User	LSTQ $\uparrow$	mAQ $\uparrow$	mIoU $\uparrow$
1	Geo-4D (Ours)	70.3	70.0	70.5
2	guysh	70.2	70.6	69.7
2	YAKD	68.6	70.8	66.5
3	hphnngcquan	68.5	70.4	66.7
4	KadirYilmaz	68.4	67.3	69.6

Appendix

Table A.1 : SemanticKITTI [ 27 ] Test Set Leaderboard (Top 5). mAQ and mIoU denote S assoc and S cls , respectively.

Method	LSTQ	S ${}_{\text{assoc}}$	S ${}_{\text{cls}}$
Hungarian Matching	72.3	77.6	67.5
Greedy Matching	72.7	78.3	67.5

Appendix

Table A.2 : Ablation on matching strategy.

		Memory (GB)	Time (s)	Validation set					Test set
Method	# Scan	Memory (GB)	Time (s)	LSTQ	S ${}_{\text{assoc}}$	S ${}_{\text{cls}}$	IoU St	IoU Th	LSTQ	S ${}_{\text{assoc}}$	S ${}_{\text{cls}}$	IoU St	IoU Th
(a) IoU-based Association
4D-PLS [ 3 ]	2	-	-	59.9	58.8	61.0	65.0	63.1	-	-	-	-	-
4D-PLS [ 3 ]	4	8.0	3.48	62.7	65.1	60.5	65.4	61.3	56.9	56.4	57.4	66.9	51.6
4D-Stop [ 4 ]	2	15.9	3.67	66.4	71.8	61.4	64.9	64.1	62.9	67.3	58.8	68.3	53.3
4D-Stop [ 4 ]	4	18.9	4.32	67.0	74.4	60.3	65.3	60.9	63.9	69.5	58.8	67.7	53.8

Appendix

Table A.3 : Analysis on performance over the number of superimposed scans.

		Assoc. Training	Validation set					Test set
Method	# Scan	Assoc. Training	LSTQ	S ${}_{\text{assoc}}$	S ${}_{\text{cls}}$	IoU St	IoU Th	LSTQ	S ${}_{\text{assoc}}$	S ${}_{\text{cls}}$	IoU St	IoU Th
DS-Net [ 32 ]	1	-	-	-	63.5	64.6	62.0	-	-	60.6	66.9	52.0
$\boldsymbol{+}$ CA-Net [ 14 ]	1	✓	67.4	71.6	63.5	64.6	62.0	62.1	63.7	60.6	66.9	52.0
$\boldsymbol{+}$ Geo-4D (Ours)	1	✗	68.5 (+1.1)	74.0 (+2.4)	63.5	64.6	62.0	62.9 (+0.8)	65.3 (+1.6)	60.6	66.9	52.0
Mask4Former ∗ [ 6 ]	1	-	-	-	67.5	65.2	70.6	-	-	70.5	72.7	67.4
$\boldsymbol{+}$ CA-Net ∗	1	✓	70.4	73.4	67.5	65.2	70.6	66.5	62.8	70.5	72.7	67.4

Appendix

Table A.4 : Comparison of Geo-4D with CA-Net using different frame-wise segmentation backbones.

	4DMatch-F				4DLoMatch-F
Method	EPE $\downarrow$	AccS $\uparrow$	AccR $\uparrow$	Outlier $\downarrow$	EPE $\downarrow$	AccS $\uparrow$	AccR $\uparrow$	Outlier $\downarrow$	Time ∗
ICP [ 42 ]	0.296	2.96	12.06	71.50	0.565	0.14	0.74	90.87	0.10
Sinkhorn [ 52 ]	0.308	2.76	8.13	79.86	0.505	0.20	0.86	89.47	3.76
NICP [ 60 ]	0.325	6.44	12.10	80.40	0.517	0.18	0.73	92.37	4.80
NSFP [ 61 ]	0.265	8.66	18.65	64.96	0.495	0.38	1.56	84.77	39.54
Ours (ICP+GGSM)	0.210	2.51	11.26	62.58	0.431	0.18	1.19	78.80	1.62

Appendix

Table A.5: Quantitative comparison on 4DMatch-F [ 59 ] and 4DLoMatch-F [ 59 ] .

Setting	LSTQ $\uparrow$	$S_{\rm assoc}\uparrow$	$S_{\rm cls}\uparrow$	Time (s) $\downarrow$
w/o DBSCAN	71.8	75.8	67.5	0.34
w/ DBSCAN	72.7	78.3	67.5	1.10

Appendix

Table A.6 : Instance refinement.

Thing classes (AQ / IoU)

Stuff classes (IoU only)

Model

LSTQ

S assoc

S cls

\blacksquare

barrier

\blacksquare

bicycle

\blacksquare

bus

\blacksquare

car

\blacksquare

const. veh.

\blacksquare

motorcycle

\blacksquare

pedestrian

\blacksquare

traffic cone

\blacksquare

trailer

\blacksquare

truck

\blacksquare

drive. surf.

\blacksquare

other flat

\blacksquare

sidewalk

\blacksquare

terrain

\blacksquare

manmade

\blacksquare

vegetation

Mask4Former [ 6 ]

67.6

63.0

72.5

48.3/76.9

42.6/37.0

60.5/90.2

76.7/92.7

42.7/44.9

61.1/77.0

53.5/74.0

48.9/55.7

38.7/47.8

62.2/77.8

95.6

69.2

71.5

74.6

88.5

87.0

Mask4D [ 12 ]

58.5

54.1

63.1

29.7/64.3

27.8/20.5

60.5/79.2

72.5/83.5

29.5/21.8

52.5/72.0

42.3/65.9

39.5/43.0

19.2/35.2

58.9/71.8

94.2

53.9

64.9

71.3

85.2

84.5

CA-Net [ 14 ]

63.4

56.6

71.0

30.1/74.5

31.9/29.1

56.5/88.5

73.3/91.2

38.6/43.2

63.2/78.7

53.6/71.2

34.6/54.6

39.5/48.3

62.1/77.3

95.6

68.4

70.3

72.7

86.7

85.2

Geo-4D (Ours)

67.9

65.0

71.0

40.7/74.5

44.1/29.1

58.9/88.5

80.9/91.2

40.1/43.2

65.9/78.7

67.2/71.2

52.0/54.6

40.6/48.3

65.7/77.3

95.6

68.4

70.3

72.7

86.7

85.2

Appendix

Table A.7 : Comparison of the quantitative results on panoptic nuScenes [ 28 ] .

Figure A.1 : Qualitative comparison on semanticKITTI [ 27 ] validation set. We visualize the association results over five consecutive scans for both the baselines and our Geo-4D . Different colors represent different instances. The dotted boxes zoom in for clear comparison. For clarity, we use circled numbers (①–⑤) to represent the frame indices.

Figure A.2 : Qualitative comparison on panoptic nuScenes [ 28 ] . We visualize the association results over five consecutive scans for both the baselines and our Geo-4D . Different colors represent different instances. The dotted boxes zoom in for clear comparison. For clarity, we use circled numbers (①–⑤) to represent the frame indices.

Explore similar work

Unsupervised Point Cloud Registration via Training-Time Semantic Guidance
Sep 14, 2026Kezheng Xiong, Shiyun Xu, Sheng Ao +3Point Cloud Registration Light Detection And Ranging Point Clouds
Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University, China · Fujian Key Laboratory of Urban Intelligent Sensing and Computing, Xiamen University, China
Not All Points Are Equal: Uncertainty-Aware 4D LiDAR Scene Synthesis
Jun 1, 2026Xiang Xu, Alan Liang, Youquan Liu +5Scene Understanding Embodied Artificial Intelligence
1NUAA · 2NUS · 3FDU +4
Self-supervised Geometry Reasoning for LiDAR Simultaneous Localization and Mapping
Jun 29, 2026Jiwoo Kim, Jinwoo Lee, Woojae Shin +2Simultaneous Localization And Mapping Point Clouds
Korea Advanced Institute of Science and Technology (KAIST), Daejeon, Republic of Korea · Daegu Gyeongbuk Institute of Science and Technology (DGIST), Daegu, Republic of Korea

Sep 14, 2026, cs.CVJ/K move · Enter open · S save

Unsupervised Point Cloud Registration via Training-Time Semantic Guidance

Kezheng Xiong, Shiyun Xu, Sheng Ao, Siqi Shen+2

Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University, China · Fujian Key Laboratory of Urban Intelligent Sensing and Computing, Xiamen University, China

Point Cloud Registration Light Detection And Ranging Point Clouds Image Registration Geometric Machine Learning Guidance

Open paper AskPDF arXiv

Unsupervised registration of large-scale LiDAR point clouds remains challenging due to the geometric ambiguity inherent in outdoor scenes, which degrades pseudo-label quality and leads to suboptimal convergence, particularly for sparse, low-resolution scans such as those from nuScenes. We reveal that registration models intrinsically encode semantic awareness that strongly correlates with registration accuracy, albeit without explicit semantic supervision. However, this native awareness is fragile: noisy supervision arising from geometric ambiguity in unsupervised settings rapidly erodes the learned semantic structure, causing performance collapse. To this end, we propose CAESAR, a teacher-student framework guided by an off-the-shelf 3D segmentation model exclusively during training. We observe that potential inlier matches are often buried just beneath a few spurious neighbors in the noisy feature space, motivating Dual-Cue Guided Re-Matching to recover them through reselection rather than simply rejecting. Building on this, a train-only Semantic-Geometric Label Mining performs lightweight, batch-specific teacher refinement and mines reliable pseudo-labels under semantic guidance. We further introduce Semantic Predictive Distillation to consolidate the student's semantic awareness in the feature space. Extensive experiments on KITTI and nuScenes demonstrate state-of-the-art performance, with pronounced gains on the challenging nuScenes benchmark. Crucially, CAESAR incurs zero inference overhead and requires no semantic annotations on the registration data. Code will be released.

Abstract

Explore similar work

Unsupervised Point Cloud Registration via Training-Time Semantic Guidance

Not All Points Are Equal: Uncertainty-Aware 4D LiDAR Scene Synthesis

Self-supervised Geometry Reasoning for LiDAR Simultaneous Localization and Mapping