DGLSS-NL: LiDAR Segmentation with Noisy Labels

• The paper introduces the DGLSS-NL framework, combining dual-view augmentation and confidence-aware filtering to tackle label noise in LiDAR semantic segmentation.
• It establishes a benchmark using synthetic symmetric noise with evaluations on SemanticKITTI, nuScenes, and SemanticPOSS via mIoU, AM, and HM metrics.
• DuNe's architecture fuses strong and weak views through PolarMix and sparse convolutions, ensuring robust performance under domain shifts and annotation imperfections.

Domain Generalization for LiDAR Semantic Segmentation under Noisy Labels (DGLSS-NL) defines a framework and benchmark for studying the simultaneous challenges of domain shift and label imperfection in 3D LiDAR semantic segmentation. In practical autonomous driving, LiDAR-based perception models are deployed across environments, vehicles, and sensors that differ significantly from the training domain. Moreover, annotation noise arises due to occlusion, sensor sparsity, and human error. Unlike noisy-label research in 2D images, the sparse and irregular nature of point clouds constrains the effectiveness of existing noisy-label learning methods. Addressing DGLSS-NL entails the development and assessment of segmentation models that can generalize to entirely new domains and maintain robustness against noisy (incorrect or ambiguous) supervisions (Kong et al., 10 Oct 2025).

1. Problem Definition and Benchmarking

DGLSS-NL formalizes the task of single-source, label-noisy domain generalization in LiDAR semantic segmentation (Kong et al., 10 Oct 2025). The setup assumes a segmentation model trained on a source domain (e.g., SemanticKITTI) under symmetric label noise; the model is then evaluated on multiple unseen target domains (e.g., nuScenes, SemanticPOSS). The benchmark employs synthetic noise (e.g., 10–50% symmetric label corruption) injected into trusted datasets to emulate real-world annotation imperfection. This dual challenge puts emphasis on both out-of-distribution generalization and robustness to supervision noise. The chosen evaluation metrics are mean Intersection over Union (mIoU), Arithmetic Mean (AM), and Harmonic Mean (HM) of mIoUs across multiple target domains.

Dataset	mIoU (10% noise)	AM	HM
SemanticKITTI	56.86%	49.57%	48.50%
nuScenes	42.28%
SemanticPOSS	52.58%

These numbers illustrate state-of-the-art robustness achieved by the DuNe framework in the DGLSS-NL setting (Kong et al., 10 Oct 2025).

2. Baseline Noisy-Label Strategies and Limitations

Prior to DGLSS-NL, noisy-label learning in images featured methods such as Twin Contrastive Learning (TCL) (Kong et al., 10 Oct 2025), Dynamic Instance Specific Selection (DISC), and Negative and Partial Learning (NPN). TCL employs Gaussian modeling and pseudo-labeling with a contrastive loss to separate clean and noisy samples. DISC dynamically identifies likely clean samples during training. NPN differentiates candidate and complementary label sets to penalize over-confident predictions on the latter. When transplanted directly to sparse 3D LiDAR data, these methods encounter obstacles: costly clustering at varying densities (TCL), expensive per-point alignment and limited stability (DISC), or insufficient adaptation to the unique redundancy and sparsity patterns of LiDAR data (NPN) (Kong et al., 10 Oct 2025).

3. The DuNe Dual-View Framework

DuNe is an architecture specifically engineered for DGLSS-NL, built upon several technical pillars (Kong et al., 10 Oct 2025):

• Dual-View Augmentation via PolarMix: Each LiDAR scan is augmented to yield a "strong" view (rich geometric perturbations, instance rotate-paste, domain-mixing) and a "weak" view (structural fidelity, sparsity simulation).
• Sparse Convolutional Backbone: Both views are processed by a shared encoder (MinkowskiEngine-based), followed by feature fusion through a lightweight decoder.
• Feature Consistency Losses:
  • Sparsity-Invariant Feature Consistency (SIFC): Enforces alignment between features extracted from the sparsity-augmented weak branch and its original.
  • Semantic Correlation Consistency (SCC): Aligns inter-class relationships via class prototypes.
  • Dual-View Feature Consistency (ℒ_FC): Encourages similar latent representations between strong and weak branches, thus regularizing the model.
• Confidence-Aware Filtering: Inspired by partial and negative supervision (NPN), the label space for each point is decomposed into a candidate set (labels predicted by both views) and a complementary set. The partial label learning (PLL) term encourages prediction of candidate labels, while the negative learning (NL) term penalizes selection from the complementary set. This explicit mechanism downweights or corrects predictions suspected of being influenced by label noise.

This architecture jointly leverages the diversity of strong augmentation and the structural reliability of weak augmentation, enabling robust learning from noisy and shifted training distributions.

4. Experimental Results and Metric Analysis

Under symmetric label noise (10%, 20%, 50% corruption), DuNe exhibits stable generalization with less performance degradation compared to conventional or naive transfer approaches:

• Mean Intersection over Union (mIoU) is computed per class and averaged:

  $$\mathrm{mIoU} = \frac{1}{C} \sum_{i=1}^C \frac{TP_i}{TP_i + FP_i + FN_i}$$

  where $C$ is the number of classes.

• Aggregate Metrics are reported for multi-domain evaluation:
  • Arithmetic Mean: $$\mathrm{AM} = \frac{1}{N} \sum_{i=1}^N \mathrm{mIoU}_i$$
  • Harmonic Mean: $$\mathrm{HM} = N / \sum_{i=1}^N (1 / \mathrm{mIoU}_i)$$

DuNe attains 56.86% mIoU on SemanticKITTI, 42.28% on nuScenes, and 52.58% on SemanticPOSS at 10% noise; corresponding AM and HM are 49.57% and 48.50%. Baseline models trained with noisy labels deteriorate sharply under increasing noise, while DuNe maintains substantially higher accuracy (Kong et al., 10 Oct 2025).

Comparison with transferred 2D noisy-label algorithms further establishes the dual-view consistency and partial/negative supervision as essential for robust LiDAR semantic segmentation in the DGLSS-NL task.

5. Comparative Analyses and Broader Algorithmic Context

While DuNe sets a clear benchmark, several algorithmic paradigms from recent literature are relevant for DGLSS-NL, especially in relation to handling sensor and domain shifts in conjunction with label imperfections:

• Completion-based Canonicalization: SVCN and "complete & label" pipelines attempt to canonicalize domain discrepancies via surface completion and local adversarial learning. This strategy can act as a denoising stage, mitigating label noise before semantic assignment (Yi et al., 2020).
• Geometric and Graph-Based Propagation: 3DLabelProp and similar methods use spatial graphs and label diffusion, which empirically demonstrate noise suppression by leveraging geometric proximity and redundant observations across sequential scans (Sanchez et al., 2022, Sanchez et al., 24 Jan 2025).
• Sample Mixing and Ensemble Pseudo-labeling: Robust pseudo-labeling (CoSMix), ensemble aggregation, and bidirectional mixing show strong promise for reducing the impact of label errors and domain-induced uncertainty (Saltori et al., 2022, Kaushik et al., 24 Jul 2025).
• Domain-Agnostic Multi-Source Learning: COLA and Point-MoE exploit label coarsening, domain-routed expert specialization, and multi-dataset fusion to encourage invariance to both domain shift and annotation noise (Sanchez et al., 2023, Chen et al., 29 May 2025).
• Noise-Aware Generalization Theory: Recent theory suggests that combining domain-wise comparison (DL4ND) with within-domain noise detection enables effective separation of structural shifts from annotation corruption—particularly valuable in sparse, multi-domain, 3D regimes (Wang et al., 3 Apr 2025).

6. Implications, Limitations, and Future Directions

DGLSS-NL advances the field by systematically evaluating methods under the composite pressure of domain shift and label noise. The proposed dual-view consistency and confidence-aware loss filtering architecture achieves state-of-the-art generalization. This paradigm brings several broader implications:

• Resilience in Real-World Autonomy: Models robust to DGLSS-NL are better poised for safe deployment in autonomous vehicles operating in unseen or evolving domains, where label and sensor variability are inevitable.
• Benchmarks and Further Research: The DGLSS-NL benchmark provides a basis for deeper investigation. Future directions include handling asymmetric and structured label noise, introducing adaptive multimodal fusion, and integrating uncertainty quantification in the context of severe domain and annotation variability (Kong et al., 10 Oct 2025).
• Limits of Transferring 2D Methods: The evidence shows that approaches designed for dense images require substantial modification to work in sparse, irregular, and high-dimensional point cloud domains.
• Expandable Modules: The dual-view paradigm is compatible with geometric fusion, pseudo-labeling ensembles, domain-mixing, and advanced foundation model adaptation, opening avenues for even more resilient generalization architectures.

In sum, DGLSS-NL and architectures such as DuNe represent a technically rigorous framework for developing and benchmarking LiDAR semantic segmentation models that are simultaneously robust to domain shifts and to noisy supervision—a critical requirement for real-world, reliable autonomous driving systems (Kong et al., 10 Oct 2025).