Exploring Single Domain Generalization of LiDAR-based Semantic Segmentation under Imperfect Labels

TL;DR

This work addresses the challenge of domain generalization for LiDAR-based 3D semantic segmentation under noisy labels. It introduces the DGLSS-NL benchmark and proposes DuNe, a dual-view learning framework that combines a strong geometry-aware view with a weak view, bottleneck feature consistency, and confidence-aware partial/negative supervision to mitigate label noise under domain shifts. Empirical results show that DuNe achieves state-of-the-art robustness across SemanticKITTI, nuScenes, and SemanticPOSS under symmetric label noise, with notable improvements in cross-domain generalization (AM and HM gains) compared to both clean-label baselines and transferred image-domain NL methods. The work provides a practical, geometry-sensitive approach for robust autonomous driving perception, and offers a reusable benchmark and code to spur further research in noise-robust LiDAR perception under domain shift.

Abstract

Accurate perception is critical for vehicle safety, with LiDAR as a key enabler in autonomous driving. To ensure robust performance across environments, sensor types, and weather conditions without costly re-annotation, domain generalization in LiDAR-based 3D semantic segmentation is essential. However, LiDAR annotations are often noisy due to sensor imperfections, occlusions, and human errors. Such noise degrades segmentation accuracy and is further amplified under domain shifts, threatening system reliability. While noisy-label learning is well-studied in images, its extension to 3D LiDAR segmentation under domain generalization remains largely unexplored, as the sparse and irregular structure of point clouds limits direct use of 2D methods. To address this gap, we introduce the novel task Domain Generalization for LiDAR Semantic Segmentation under Noisy Labels (DGLSS-NL) and establish the first benchmark by adapting three representative noisy-label learning strategies from image classification to 3D segmentation. However, we find that existing noisy-label learning approaches adapt poorly to LiDAR data. We therefore propose DuNe, a dual-view framework with strong and weak branches that enforce feature-level consistency and apply cross-entropy loss based on confidence-aware filtering of predictions. Our approach shows state-of-the-art performance by achieving 56.86% mIoU on SemanticKITTI, 42.28% on nuScenes, and 52.58% on SemanticPOSS under 10% symmetric label noise, with an overall Arithmetic Mean (AM) of 49.57% and Harmonic Mean (HM) of 48.50%, thereby demonstrating robust domain generalization in DGLSS-NL tasks. The code is available on our project page.

Figures (3)

• Figure 1: We inject symmetric label noise into the training set according to predefined noise ratios. The figure reports the segmentation performance on the test set, where the vertical axis shows mIoU(%)($\uparrow$), the horizontal axis corresponds to the applied methods, and the third dimension indicates the results under different ratios of symmetric noise. DuNe consistently achieves the best results across all noise levels.
• Figure 2: Overview of our proposed dual-view training pipeline - DuNe. Each input LiDAR scan is first augmented by PolarMix xiao2022polarmix to generate a strong view and a weak view. Both views are further processed with sparsity augmentation. For the strong view, the entire augmented scan is used to construct candidate labels and complementary labels bib:npn, which provide noise-robust supervision. For the weak view, we explicitly split the original and sparsity-augmented versions to form paired inputs, and enforce consistency loss and semantic correlation loss between them. This design allows the framework to jointly exploit label-level robustness and view-level consistency for improved noisy-label learning.
• Figure 3: Qualitative results under 10%, 20%, and 50% label noise. For each noise ratio, we compare segmentation predictions of our framework DuNe with the baseline DGLSS. Corresponding class-wise prediction histograms (10 categories) are also shown. The left column visualizes predictions, where circles highlight differences between methods. The right column presents class-wise prediction distributions: colored bars indicate results from our method, while gray bars correspond to predictions from the DGLSS baseline. Our method produces more coherent segmentations and maintains balanced class distributions, even under severe noise.