LiDAR-based 3D Object Detection (L3OD) is a fundamental 3D perception task, yet its performance is often compromised by domain shifts arising from diverse sensor configurations and environmental conditions. Existing Test-Time Adaptation (TTA) methods primarily leverage self-supervision in the spatial domain, yet they frequently confound semantic content with domain style due to the sparse, non-Euclidean nature of point clouds. To overcome this limitation, we propose FIND (Frequency INvariance Disentanglement), a novel TTA framework that precisely disentangles domain-invariant content from domain-specific interference within the frequency domain. Our methodology utilizes B-Spline fitting to construct locally adaptive filters integrated into a dual-stream architecture: an Invariant Stream extracts robust features to guide pseudo-labeling, while a Specific Stream enforces consistency against domain perturbations. By combining invariance-driven alignment with stability-guided regularization, our approach dynamically extracts robust domain-invariant features while suppressing domain-specific interference. Extensive experiments on cross-dataset adaptation and robust corruption benchmarks demonstrate that FIND significantly outperforms state-of-the-art methods. The source code is publicly available at https://github.com/llldx03/FIND.
Figures & tables
Figure 1: (a) Illustration of domain shift types in L3OD. (b) Spatial constraints of point clouds, i.e., complex entanglement. (c) Disentanglement paradigm through Fourier transformation.
Figure 2: Spectral entanglement analysis of 3D point clouds. (a) shows compromised geometric integrity resulting from rigid filtering distinct frequency bands. (b) records cross-domain KL divergence ( DKL ) across 12 uniform frequency bins. (c) investigates Spearman correlation between Coefficient of Variation ( Cv ) and KL divergence within each bin for both source and target domains. Experiments are conducted on nuScenes → KITTI scenario. More empirical analysis on other scenarios are provided in Appendix B .
Figure 3: Overview of our FIND method. (a) GFT foundation for spectral analysis of point cloud. (b) Dual-stream architecture with distinct filter modules to disentangle domain-invariant semantic content (Sec. 5.1 ). (c) FIND employs B-Spline frequency filters to capture narrow-band invariant and specific components. (d) Through the disentangled learning strategy (Sec. 5.2 ), L3OD model and filtering modules can be optimized jointly to foster a virtuous cycle.
Method
Reference
nuScenes → KITTI
Waymo → KITTI
APBEV/AP3D
Closed Gap
APBEV/AP3D
Closed Gap
Source
-
45.50 / 13.78
-
65.64 / 26.74
-
Oracle
-
83.29 / 73.45
-
83.29 / 73.45
-
Tent
ICLR’21
45.13 / 19.33
− 0.98% / + 9.30%
68.54 / 34.57
+ 16.43% / + 16.76%
CoTTA
CVPR’22
68.81 / 47.61
+ 61.68% / + 56.70%
67.46 / 35.34
+ 10.31% / + 18.41%
Reg-TTA3D
ECCV’24
70.21 / 49.32
+ 65.39% / + 59.56%
81.60 / 61.03
+ 90.42% / + 73.41%
Table 1: Comparisons with state-of-the-art methods on nuScenes → KITTI and Waymo → KITTI scenarios. We report APBEV/AP3D of Car category in moderate difficulty. "Source" and "Oracle" denote model performance with no adaptation and fully supervised training, respectively. We highlight the best and second results with bold and underline respectively. Due to the space limit, the complete results with more scenarios and categories are provided in Tab. 7 of Appendix. C .
Method
Reference
Weather
Disturbance
Sensor
Avg.
Fog
Rain
Snow
Motion.
Beam.
Cross T.
Inc.
Cross S.
Source
-
67.72
75.99
59.32
38.77
53.85
75.21
27.49
41.24
54.95
Tent
ICLR’21
68.84
76.22
59.69
38.35
53.76
75.10
28.49
41.35
55.23
CoTTA
CVPR’22
68.45
76.40
59.49
38.82
54.08
75.37
30.56
42.25
55.68
Reg-TTA3D
ECCV’24
68.00
77.51
60.72
40.34
56.23
76.01
32.17
44.02
56.88
MOS
ICLR’25
69.11
77.79
62.72
40.59
55.91
75.47
34.53
43.68
57.48
Table 2: Comparisons with state-of-the-art methods on KITTI → KITTI-C (heavy severity). We report AP3D of Car category in moderate difficulty. The complete results of moderate and easy severity levels are provided in Tab. 8 - 9 of Appendix. C .
Method
Reference
Weather
Disturbance
Sensor
Avg.
Fog
Rain
Snow
Motion.
Beam.
Cross T.
Inc.
Cross S.
Source
-
16.97
26.19
11.09
10.23
12.72
20.31
3.55
14.22
14.41
Tent
ICLR’21
23.28
39.05
23.55
7.38
19.37
28.21
9.61
13.44
20.49
CoTTA
CVPR’22
31.39
50.10
25.11
18.99
27.00
35.19
10.10
13.26
26.39
Reg-TTA3D
ECCV’24
37.37
59.16
27.90
23.16
32.13
41.35
13.80
19.99
31.86
MOS
ICLR’25
43.26
58.32
28.87
24.50
32.20
39.63
14.96
19.19
32.62
Table 3: Comparisons with state-of-the-art methods on Waymo → KITTI-C (heavy level).
Src.
Lcons
Lst
Lcomp
nuScenes → KITTI
Waymo → KITTI
APBEV
AP3D
APBEV
AP3D
✔
45.50
13.78
65.64
26.74
✔
69.32
48.81
72.57
58.93
✔
60.84
35.87
69.88
43.91
✔
✔
68.71
49.76
73.95
60.62
✔
✔
72.56
52.49
78.27
64.48
Table 4: Effects of components in our method. We conduct ablation studies of the invariance-driven alignment loss Lcons , stability-guided regularization loss Lst and complementary constraint loss Lcomp on cross-dataset scenarios. "Src." denotes the source model without adaptation.
Method
Time (ms)
Memory (MiB)
Tent
88 ± 5
7,476
CoTTA
179 ± 10
10,462
Reg-TTA3D
206 ± 14
8,564
MOS
3088 ± 160
14,134
DUO
158 ± 8
8,312
Ours ( FIND )
194 ± 18
11,540
Table 5: Computational efficiency on Waymo → KITTI-C using a single NVIDIA RTX 4090 with batch size 8. Time records the average processing time per point cloud in KITTI. Memory records max CUDA memory allocated during adaptation.
Figure 4: (a) Performance with varying numbers of bins N . (b) Performance with different B-Spline order K . Experiments are conducted on nuScenes → KITTI.
Figure 5: Visualization of point clouds in KITTI before and after filtering. Left: Raw data entangled with sensor noise. Right: Output from adaptive filters exhibiting geometric structure.
Appendix figures & tables14 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 6: Controlled sensor-only empirical analysis on KITTI → KITTI-C. The underlying scene geometry is fixed while only sensor/weather conditions vary. (a) Average spectral energy μ(E) span all bands regardless of corruption type. (b) Per-band DKL fluctuation pattern is consistent under different corruptions. (c) Cv profiles remain stable across different corruption types.
Figure 7: Empirical analysis of KL Divergence and Coefficient of Variation on Waymo→KITTI and KITTI→KITTI-C (fog, heavy) scenarios.
Scenario
Source Domain
Target Domain
ρ
p -value
ρ
p -value
nuScenes → KITTI
0.97
1.75×10−7
0.97
1.75×10−7
Waymo → KITTI
0.96
7.80×10−7
0.96
1.18×10−6
KITTI → KITTI-C (fog)
0.95
2.04×10−6
0.94
3.93×10−6
Appendix
Table 6: Spearman rank correlation ( ρ ) between Cv and DKL across scenarios. All correlations are statistically significant ( p<10−5 ).
Method
Reference
nuScenes → KITTI
Waymo → KITTI
Waymo → nuScenes
Car
Ped.
Cyc.
Car
Ped.
Cyc.
Car
Ped.
Cyc.
Source
-
13.78
27.75
8.37
26.74
36.01
38.29
17.22
0.78
1.03
Oracle
-
73.45
41.33
60.32
73.45
41.33
60.32
34.87
18.92
11.73
Tent
ICLR’21
19.33
29.55
11.12
34.57
38.39
41.56
17.83
1.78
0.92
CoTTA
CVPR’22
47.61
33.45
12.46
35.34
42.05
44.75
18.12
1.73
2.78
Reg-TTA3D
ECCV’24
49.32
33.63
16.44
61.03
45.53
49.77
18.26
2.23
4.73
Appendix
Table 7: Comparisons with state-of-the-art methods on nuScenes → KITTI, Waymo → KITTI and Waymo → nuScenes scenarios. We record AP3D of Car, Pedestrian (Ped.), Cyclist (Cyc.) categories in moderate difficulty. We highlight the best and second results with bold and underline respectively.
Method
Reference
Weather
Disturbance
Sensor
Avg.
Fog
Rain
Snow
Motion.
Beam.
Cross T.
Inc.
Cross S.
Source
-
72.36
79.13
64.09
39.51
57.53
78.26
30.39
47.21
58.56
TENT
ICLR’21
73.07
78.98
64.18
39.66
58.01
78.19
30.50
47.34
58.74
CoTTA
CVPR’22
73.11
80.59
65.93
40.51
59.09
78.72
35.34
48.36
60.21
REG-TTA3D
ECCV’24
73.59
81.93
67.67
41.37
60.26
79.06
39.51
49.40
61.60
MOS
ICLR’25
74.02
81.91
67.87
41.63
60.48
79.10
40.29
49.28
61.82
Appendix
Table 8: Comparisons with state-of-the-art methods on KITTI → KITTI-C (moderate severity). We record AP3D of Car category in moderate difficulty and highlight the best and second results with bold and underline respectively.
Method
Reference
Weather
Disturbance
Sensor
Avg.
Fog
Rain
Snow
Motion.
Beam.
Cross T.
Inc.
Cross S.
Source
-
84.51
88.37
71.29
43.49
75.38
88.10
48.11
68.02
70.91
TENT
ICLR’21
84.21
88.68
71.54
43.88
76.05
88.30
50.16
69.06
71.49
CoTTA
CVPR’22
84.89
88.54
73.53
44.46
76.98
88.70
53.66
69.88
72.58
REG-TTA3D
ECCV’24
85.03
88.91
74.79
44.58
78.27
88.90
58.39
70.68
73.69
MOS
ICLR’25
85.22
89.32
75.07
44.70
78.74
89.22
59.13
71.87
74.16
Appendix
Table 9: Comparisons with state-of-the-art methods on KITTI → KITTI-C (easy severity). We record AP3D of Car category in moderate difficulty and highlight the best and second results with bold and underline respectively.
Stage
Time (ms)
Proportion
Bounding-box prediction
73
37.63%
Semantic-aware graph construction
6
3.09%
B-Spline spectral filtering
27
13.92%
Dual-stream disentangled adaptation
88
45.36%
Full FIND
194
100.00%
Appendix
Table 10: Processing-time breakdown of FIND under the same measurement setting. Each stage’s time is normalized by the number of processed frames.
Method
Reference
nuScenes → KITTI
Waymo → KITTI
KITTI → KITTI-C
Source
-
13.78
26.74
54.95
SC-NNK
ICASSP’23
42.69
52.37
55.46
L-TTA
NeurIPS’24
44.57
51.33
56.25
SPA
ICML’25
48.19
57.56
55.92
Ours ( FIND )
-
54.66
67.09
59.62
Appendix
Table 11: Comparisons with point-cloud denoising and frequency-domain TTA methods. We report AP3D of the Car category at moderate detection difficulty. KITTI → KITTI-C results are averaged over eight corruption types at heavy severity. We highlight the best and second results with bold and underline , respectively.
Method
Protocol
Fog
Rain
Snow
Avg.
Source
Frozen
67.72
75.99
59.32
67.68
MOS
Independent
69.11
77.79
62.72
69.87
MOS
Continual
69.11
76.53
60.85
68.83
Ours ( FIND )
Independent
71.27
78.23
64.89
71.46
Ours ( FIND )
Continual
71.27
78.04
63.96
71.09
Appendix
Table 12: Continual adaptation under Fog → Rain → Snow shifts on KITTI → KITTI-C at heavy severity. We report AP3D of the Car category at moderate detection difficulty. “Independent” denotes separate adaptation to each corruption, while “Continual” denotes adaptation along the stream without resets or domain labels. We highlight the best results within each adaptation protocol with bold .
Figure 8: (a) Performance with varying weights βst of Lst and βcomp of Lcomp . (b) Performance with different decay scale factor τ for stability mask. (c) Performance with distinct threshold initializations of Invariant Stream.
Filter
nuScenes → KITTI
Waymo → KITTI
APBEV/AP3D
Closed Gap
APBEV/AP3D
Closed Gap
Source
45.50 / 13.78
-
65.64 / 26.74
-
Rigid
54.31 / 32.56
+ 23.31% / + 31.47%
66.31 / 39.10
+ 3.80% / + 26.46%
Global
69.72 / 48.38
+ 64.09% / + 57.99%
72.55 / 50.21
+ 39.15% / + 50.25%
Ours ( B-Spline )
75.21 / 54.66
+ 78.62% / + 68.51%
82.35 / 67.09
+ 94.67% / + 86.38%
Appendix
Table 13: Comparisons with different filter mechanisms on nuScenes → KITTI and Waymo → KITTI scenarios. We report APBEV/AP3D of Car category in moderate difficulty. "Source" denotes model performance with no filters or adaptation. "Rigid" and "Global" denote model performance with rigid band-pass filters and polynomial filters under global control, respectively. We highlight the best results with bold .
Figure 9: Visualization of B-Spline filter response functions within Invariant Stream under 4 cases: N=4,N=18,K=1,K=6 . The response magnitudes range from 0 to 1, representing filtering strength, where higher values indicate greater feature preservation.
Figure 10: Visualization of adapted B-Spline filter response functions within Invariant Stream and Specific Stream. The response magnitudes range from 0 to 1, representing filtering strength, where higher values indicate greater feature preservation.
Method
nuScenes → KITTI
Waymo → KITTI-C
AP3D
Gap
AP3D
Gap
MOS
51.11
− 3.55
32.62
− 1.37
Ours ( FIND )
54.66
0.00
33.99
0.00
+ Localization noise
54.22
− 0.44
33.13
− 0.86
+ Missed boxes
52.08
− 2.58
32.39
− 1.60
+ False boxes
54.30
− 0.36
33.46
− 0.53
Appendix
Table 14: Robustness to errors in predicted boxes used for graph construction. We report AP3D of the Car category at moderate detection difficulty; Waymo → KITTI-C is averaged over eight corruption types at heavy severity. Gap denotes the difference in AP3D relative to FIND without added perturbations. We highlight the best results with bold .
Lidar-based 3D object detection locates and classifies objects in point clouds, supporting perception for autonomous vehicles, robots, and other mobile systems.
In autonomous driving, 3D LiDAR semantic segmentation is expected to adapt to evolving environments and recognize previously unseen categories, making generalized category discovery under test-time adaptation (GCD-TTA) an important problem. However, existing works treat GCD and TTA separately, while the limited semantic information in 3D point clouds makes GCD-TTA under domain shift more prone to feature collapse when combining GCD with TTA. To address this issue, we propose a two-stage framework for GCD-TTA in 3D LiDAR semantic segmentation. Specifically, in the first stage, we adaptively aggregate local 3D features according to geometric density variations and obtain reliable pseudo labels under a global class-wise prototype consistency constraint. In the second stage, we cast category discovery as an optimal transport assignment process for globally consistent grouping of novel categories. Meanwhile, semantic features from 2D vision foundation models are fused with 3D point cloud features to enrich 3D semantic information and promote a more separable feature space. Extensive experiments on Synth4D to SemanticKITTI, SynLiDAR to SemanticKITTI, and Synth4D to nuScenes demonstrate the effectiveness of our framework. Under a unified evaluation protocol, our method achieves the best overall mIoU on 9 out of 10 benchmark splits and improves the overall mIoU by up to 12.21 points over the strongest re-implemented baseline.
Hanqi Chen, Chaoqi Chen, Tian Ji +1
CSSE, Shenzhen University Shenzhen, Guangdong, China · The Hong Kong University of Science and Technology (Guangzhou) Guangzhou, Guangdong, China
LiDAR semantic segmentation often degrades under real-world deployment due to evolving sensing conditions, while collecting new annotations for retraining is impractical. Test-time adaptation (TTA) updates model parameters online using pseudo-label supervision, but directly applying standard TTA strategies to LiDAR data is challenging. Because pseudo-label reliability is spatially heteroscedastic under range-dependent sparsity and occlusion, uniform updates on globally shared parameters can inject unstable gradients and destabilize adaptation. We propose a geometry-constrained test-time prompt tuning framework for LiDAR semantic segmentation. Our method estimates per-location sensing reliability from depth-consistent beam terminations and neighborhood support, and uses it to reweight spatial supervision. Adaptation is confined to lightweight prompt adapters inserted into a frozen backbone, with spatial gating to prevent unreliable regions from perturbing globally shared representations. A temporally smoothed prototype alignment strategy further stabilizes online updates by accumulating reliable semantic evidence over time. Experiments on standard LiDAR benchmarks demonstrate improved adaptation stability and segmentation performance under deployment variations without additional annotations.
Linlian Jiang, Wentao Ju, Sadman Rakib Pinon +4
Concordia University · Mila – Quebec AI Institute · University of Toronto
Changes in sensor height and viewpoint alter object-level point distributions, making cross-platform LiDAR unsupervised domain adaptation (UDA) difficult. Self-training uses labeled source scans and unlabeled target scans, yet a retained prediction may provide a useful target location while enclosing sparse foreground returns, background clutter, or points inconsistent with the predicted box. We refer to this mismatch as box-point inconsistency. We introduce SimFuse3D, which preserves the target placement and repairs the associated pseudo object using measured geometry from labeled source scans. Object Memory retrieves a similar labeled source instance. Target Simulation places the retrieved source geometry at the target location, aligns its points with the target viewing geometry, and filters the aligned crop to approximate the target observation. Confidence-Guided Multi-Stage Localization Reweighting (CMLR) maps each target pseudo-object confidence score to a bounded weight shared by RPN localization and R-CNN box regression. All components operate only during adaptation, leaving the detector architecture and inference graph unchanged. Across six cross-platform transfers, SimFuse3D consistently outperforms Pi3DET-Net and achieves the best performance among the compared adaptation methods on nearly all metrics. On nuScenes-to-KITTI, it ranks first among the compared adaptation methods with both evaluated detectors.
Yongchun Lin, Xinliang Zhang, Yun Zou +7
Guangdong University of Technology · SenseTime · The Hong Kong University of Science and Technology (Guangzhou)