Adverse weather conditions significantly degrade the performance of LiDAR point cloud semantic segmentation networks by introducing large distribution shifts. Existing augmentation-based methods attempt to enhance robustness by simulating weather interference during training. However, they struggle to fully exploit the potential of augmentations due to the trade-off between minor and aggressive augmentations. To address this, we propose A3Point, an adaptive augmentation-aware latent learning framework that effectively utilizes a diverse range of augmentations while mitigating the semantic shift, which refers to the change in the semantic meaning caused by augmentations. A3Point consists of two key components: semantic confusion prior (SCP) latent learning, which captures the model's inherent semantic confusion information, and semantic shift region (SSR) localization, which decouples semantic confusion and semantic shift, enabling adaptive optimization strategies for different disturbance levels. Extensive experiments on multiple standard generalized LiDAR segmentation benchmarks under adverse weather demonstrate the effectiveness of our method, setting new state-of-the-art results.
Figures & tables
Figure 1: (a) Point cloud distortions caused by adverse weather conditions. (b) Augmentation at different levels (light, moderate, and heavy), where we adjust the drop ratio (DR) for point drop and jitter std (JS) for random jittering. (c) Visualization of segmentation accuracy under different augmentation levels, showing that aggressive distortions lead to significant performance degradation.
Figure 2: Demonstration of semantic confusion and semantic shift . (a) Confusion matrices from source (normal weather) and target (adverse weather) domains. Despite domain shifts, semantic confusion remains consistent. (b) Aggressive augmentations alter point cloud density and shape, leading to semantic misalignment (e.g., car → veg. ).
Figure 3: Pipeline of A3Point. We explore an abundant augmentation space (Sec. 3.4 ) and propose two key components: SCP latent learning to capture inherent semantic confusion (Sec. 3.5 ) and SSR localization to decouple semantic shift (Sec. 3.6 ).
Methods
car
bi.cle
mt.cle
truck
oth-v.
pers.
bi.clst
mt.clst
road
parki.
sidew.
other-g.
build.
fence
veget.
trunk
terr.
pole
traf.
D-fog
L-fog
Rain
Snow
mIoU
gain
Oracle
89.4
42.1
0.0
59.9
61.2
69.6
39.0
0.0
82.2
21.5
58.2
45.6
86.1
63.6
80.2
52.0
77.6
50.1
61.7
51.9
54.6
57.9
53.7
54.7
-
Baseline
67.1
5.0
28.1
38.5
14.6
45.8
8.3
13.8
40.1
16.1
26.1
3.3
71.6
52.7
53.8
33.9
39.2
25.3
12.7
30.7
30.1
29.7
25.3
31.4
+0.0
PointDR ∗ ( Xiao et al., 2023 )
69.2
1.0
8.9
41.9
7.6
48.9
17.0
36.2
57.8
15.9
32.3
4.0
75.7
46.4
54.0
36.2
43.9
23.7
24.2
37.3
33.5
35.5
26.9
33.9
+2.5
DGUIL ( Kim et al., 2023 )
78.2
2.5
33.0
29.7
6.1
49.8
0.8
40.9
67.3
7.2
38.0
2.2
79.8
54.4
64.1
36.8
52.3
31.0
40.0
36.3
34.5
35.5
33.3
37.6
+6.2
WADG ( Du et al., 2024 )
72.0
0.0
32.9
37.0
1.9
37.7
6.8
52.9
59.9
10.7
31.8
2.2
76.0
48.8
62.7
34.0
49.3
23.6
20.4
39.5
32.5
31.7
29.4
34.8
+3.4
DGLSS ( He et al., 2024 )
69.6
0.8
42.8
34.4
8.9
41.9
12.8
44.5
52.0
14.5
30.8
6.0
77.8
51.1
57.6
38.9
43.2
29.7
30.6
34.2
34.8
36.2
32.1
36.2
+4.8
Table 1: Comparison results of [A] → [C]. ∗ denotes the reproduced result with the same backbone.
Methods
car
bi.cle
mt.cle
truck
oth-v.
pers.
bi.clst
mt.clst
road
parki.
sidew.
other-g.
build.
fence
veget.
trunk
terr.
pole
traf.
D-fog
L-fog
Rain
Snow
mIoU
gain
Oracle
89.4
42.1
0.0
59.9
61.2
69.6
39.0
0.0
82.2
21.5
58.2
45.6
86.1
63.6
80.2
52.0
77.6
50.1
61.7
51.9
54.6
57.9
53.7
54.7
-
Baseline
33.8
1.7
3.3
15.5
0.2
25.5
1.6
3.4
15.3
9.2
16.8
0.1
33.4
21.9
39.5
18.7
44.0
8.8
0.8
15.2
16.0
16.8
12.8
15.5
+0.0
PointDR ∗ ( Xiao et al., 2023 )
41.1
2.8
3.4
18.1
0.2
31.3
2.8
3.3
34.4
10.2
19.7
1.0
52.7
22.0
48.5
21.3
38.3
19.2
5.6
19.1
20.3
25.3
19.0
19.8
+4.3
WADG ( Du et al., 2024 )
33.8
1.1
2.9
17.0
0.2
26.8
1.0
4.3
53.9
5.0
20.6
2.2
64.3
27.1
53.8
27.0
37.0
28.6
8.6
21.6
23.4
27.2
21.4
21.9
+6.4
LiDARWeather ( Park et al., 2024 )
39.3
2.9
0.9
19.4
0.8
27.7
2.2
3.8
42.5
9.4
21.6
0.3
51.9
33.5
47.4
23.1
33.3
23.2
6.8
19.0
21.2
23.1
17.3
20.5
+5.0
NTN ( Park et al., 2025 )
48.4
1.5
2.4
19.4
0.2
29.1
3.2
8.9
43.5
6.7
20.5
0.0
52.2
30.1
49.8
20.0
32.9
24.7
7.5
-
-
-
-
21.1
+5.6
Table 2: Comparison results of [B] → [C]. ∗ denotes the reproduced result with the same backbone.
Figure 6Table 7Table 8
Appendix figures & tables22 assets
Supplementary material from the paper’s appendix.
Appendix
Augmentation Level
[A] → [C]
[B] → [C]
[A] → [A]
Light
41.3
27.2
62.5
Moderate
40.1
26.0
60.3
Heavy
37.5
24.5
55.8
Appendix
Table 7: Impact of different augmentation levels on domain generalization performance.
k
D
mIoU (%)
16
64
39.9
32
64
41.3
64
64
41.4
32
32
40.8
32
64
41.3
32
128
41.2
Appendix
Table 8: Ablation study on latent space dimensions.
t=2
t=3
t=4
mIoU (%)
41.0
41.3
40.9
Appendix
Table 9: Ablation study on SSR localization threshold.
λ=0.02
λ=0.1
λ=0.5
mIoU (%)
40.5
41.3
40.9
Appendix
Table 10: Ablation study on loss weighting coefficient.
Figure 6: Qualitative results on comparison between error map with semantic shift region.
Figure 7: Qualitative results on [A] → [C], where improvements are marked with boxes.
Figure 8: Agreement with EMA teacher within MSSR_eval .
Figure 9: Example scenes showing “high-distortion” subregions.
Method
Car
Road
Sidewalk
Vegetation
Terrain
Overall mIoU
EAS
61.6
50.7
21.4
47.5
35.7
27.1
A3Point
76.7
57.6
26.8
51.5
37.6
34.7
Appendix
Table 11: Performance within high-distortion subregions.
Method
GPU Memory (MB)
Iterations/s (it/s)
Total Time (h)
Baseline
8,496
2.72
24.4
PointDR
9,624
1.97
33.7
EAS
8,932
2.05
32.4
A3Point (ours)
17,144
1.61
41.3
Appendix
Table 12: Training-time overhead comparison under identical hardware and settings.
Setting
[A] → [C]
[B] → [C]
None
31.4
15.5
EAS
38.7
24.9
Prototype-based
40.0
25.4
VQ-VAE-based (ours)
41.3
27.2
Appendix
Table 13: Comparison of class-wise K-prototype versus VQ-VAE prior under identical settings.
Method
[A] → [C]
[B] → [C]
Baseline
31.4
15.5
LiDARWeather*
37.8
20.4
A3Point (ours)
40.5
25.9
Appendix
Table 14: 15-epoch (LiDARWeather-style) training. * denotes the reproduced result.
Method
[A] → [C]
[B] → [C]
Baseline*
32.3
16.1
LiDARWeather*
38.6
21.6
A3Point (ours)
41.3
27.2
Appendix
Table 15: 50-epoch training. * denotes the reproduced result.
Figure 10: Confusion matrices on SemanticSTF. Left: Oracle. Right: A3Point.
Strategy
[A] → [C]
[B] → [C]
None
38.7
24.9
Offline
40.7
26.4
Offline (ensemble)
41.0
26.9
Online (ours)
41.3
27.2
Appendix
Table 16: Comparison of offline and online confusion priors. Online (ours) consistently yields the best mIoU by maintaining alignment with the evolving decision boundary.
Run
car
bi.cle
mt.cle
truck
oth-v.
pers.
bi.clst
mt.clst
road
parki.
sidew.
other-g.
build.
fence
veget.
trunk
terr.
pole
traf.
mIoU
1
89.4
42.1
0.0
59.9
61.2
69.6
39.0
0.0
82.2
21.5
58.2
45.6
86.1
63.6
80.2
52.0
77.6
50.1
61.7
54.7
2
89.6
37.4
9.6
56.4
64.3
67.3
44.8
0.8
81.8
27.9
54.6
50.1
86.5
66.2
78.3
56.3
75.4
45.8
58.6
55.3
3
89.7
41.8
4.0
58.6
62.8
66.1
40.7
0.0
82.0
23.4
51.8
47.2
86.7
64.9
78.7
59.1
75.0
49.0
57.0
54.6
4
89.6
35.3
0.2
60.8
64.3
67.4
44.4
5.4
81.3
27.4
54.6
50.2
87.2
66.3
77.8
61.1
74.0
51.8
59.5
55.7
5
89.7
39.9
6.8
58.1
62.1
61.6
48.5
11.1
76.4
21.7
51.4
43.5
83.1
61.3
77.2
63.4
74.9
55.3
56.4
54.9
Appendix
Table 17: Oracle on [C]: per-class IoU across five independent runs. Rare classes (motorcycle, motorcyclist) exhibit high variance; 0.0 IoU appears in some runs but not others.
Method
None
light
moderate
heavy
random (light–heavy)
excessive
random* (none–excessive)
Baseline
31.4
37.6
38.0
37.1
38.7
5.1
36.5
A3Point
31.8
38.5
40.4
40.5
41.3
7.8
41.2
Appendix
Table 18: Stress test with enlarged augmentation ranges on [A] → [C]. Random: sampled within light–heavy; Random*: sampled within none–excessive.
Strategy
[A] → [C]
[B] → [C]
None
38.7
24.9
A3Point
41.3
27.2
A3Point+ (curriculum)
41.5
27.8
Appendix
Table 19: Iterative, curriculum-style A3Point that expands the augmentation ceiling in stages (light → moderate → heavy).
Method
[A] → [C]
[B] → [C]
Baseline
31.4
15.5
Baseline + EAS
38.7
24.9
PointDR*
33.9
19.8
PointDR* + EAS
39.1
25.1
LiDARWeather*
37.8
20.4
LiDARWeather* + EAS
39.6
25.9
Appendix
Table 20: Re-running prior methods under the same EAS upper bounds. * denotes the reproduced result.
Figure 11: t-SNE of the learned VQ-VAE codebook. Colors denote class-specific sub-codebooks; marker size encodes per-code variance.
Method
[A] → [C]
[B] → [C]
Baseline
31.4
15.5
EAS
38.7
24.9
EAS + MT
39.0
25.5
EAS + SCE
38.8
24.6
A3Point (ours)
41.3
27.2
Appendix
Table 21: Comparisons with other DG methods under the same augmentation space.
Variant
[A] → [C]
[B] → [C]
Baseline
31.4
15.5
EAS
38.7
24.9
Class-conditional nearest code (SSR)
38.8
24.5
Soft CE on logits with temperature (SSR)
39.7
25.8
Consistency with pre-augmentation prediction
37.2
24.4
A3Point (global nearest code in latent space)
41.3
27.2
Appendix
Table 22: Controlled comparisons of SSR supervision choices under the same augmentation space. Global nearest code (class-agnostic latent alignment) performs best.
3D point cloud segmentation assigns semantic, instance, or part labels to points for scene understanding, mapping, robotics, and scientific measurement.
Annotating large-scale LiDAR point clouds for 3D semantic segmentation is costly and time-consuming, which motivates the use of semi-supervised learning (SemiSL). Standard LiDAR SemiSL methods typically adopt a two-step training paradigm, where pseudo-labels are separately generated from a single distillation source, either from the same or another LiDAR representation. Such supervision relies on a unique source of pseudo-labels, which can reinforce confirmation bias and propagate errors during training, ultimately limiting performance. To address this challenge, we introduce CoLLiS, a novel framework that leverages Collaborative Learning for LiDAR Semi-supervised segmentation. Unlike prior paradigms with decoupled pseudo-labeling and training phases, CoLLiS trains multiple representations collaboratively in a single step by treating them as coequal students. Each student is adaptively distilled from multiple representations, while inter-student disparities are monitored online to resolve contradictory supervision and effectively mitigate confirmation bias. Extensive experiments on three datasets demonstrate that CoLLiS consistently outperforms state-of-the-art LiDAR SemiSL methods, with particularly strong gains in low-label regimes.
Bin Yang, Alexandru Paul Condurache
Bosch Research, Robert Bosch GmbH, Stuttgart, Germany · Institute for Neuro- and Bioinformatics, University of Lübeck, Lübeck, Germany
This paper studies LiDAR-assisted RGB semantic segmentation, where LiDAR point-cloud features improve dense pixel-wise predictions in the RGB image plane rather than 3D point labels or detections. Existing 2D fusion methods often project point clouds into shallow depth/range maps and apply spatially uniform fusion, discarding rich 3D structure and ignoring the range-dependent reliability of outdoor LiDAR. We propose RAHF-Net, a Range-Aware Heterogeneous Fusion Network. Its Cylindrical-to-Perspective (C2P) module projects pretrained cylindrical voxel features into the image plane, preserving structured 3D semantic priors. A Range-Aware Heterogeneous Fusion (RAHF) module then uses a soft depth-derived range mask to perform LiDAR-guided geometric enhancement in reliable near-field regions and visual-context enhancement in sparse far-field regions. On KITTI-360, RAHF-Net achieves 65.76 mIoU, outperforming representative RGB-only and RGB-LiDAR baselines, including CMX by 1.45 points. Experiments on DELIVER further demonstrate improved robustness under adverse weather, illumination degradation, and sensor corruption.
R. D. Lin, Pengcheng Weng, Ge Wang +2
School of Software Engineering, Xi’an Jiaotong University, China · School of Computer Science and Technology, Xi’an Jiaotong University, China
Semantic segmentation of 3D LiDAR point clouds is important in urban remote sensing for understanding real-world street environments. This task, by projecting LiDAR point clouds and 3D semantic labels as sparse maps, can be reformulated as a 2D problem. However, the intrinsic sparsity of the projected LiDAR and label maps can result in sparse and inaccurate intermediate 2D semantic predictions, which in return limits the final 3D accuracy. To address this issue, we enhance this task by shaping dense and accurate 2D predictions. Specifically, we develop a multi-modal segmentation model, MM2D3D. By leveraging camera images as auxiliary data, we introduce cross-modal guided filtering to overcome label map sparsity by constraining intermediate 2D semantic predictions with dense semantic relations derived from the camera images; and we introduce dynamic cross pseudo supervision to overcome LiDAR map sparsity by encouraging the 2D predictions to emulate the dense distribution of the semantic predictions from the camera images. Experiments show that our techniques enable our model to achieve intermediate 2D semantic predictions with dense distribution and higher accuracy, which effectively enhances the final 3D accuracy. Comparisons with previous methods demonstrate our superior performance in both 2D and 3D spaces.
Xiaoyu Dong, Tiankui Xian, Wanshui Gan +1
The University of Tokyo, Tokyo, Japan · RIKEN AIP, Tokyo, Japan