UniMix: Towards Domain Adaptive and Generalizable LiDAR Semantic Segmentation in Adverse Weather

TL;DR

UniMix tackles the challenge of LiDAR semantic segmentation under adverse weather by introducing a Bridge Domain via physically valid weather simulation and a Universal Mixing operator that blends samples across spatial, intensity, and semantic dimensions. The method adopts a two-stage teacher-student framework to enable both unsupervised domain adaptation and domain generalization, learning weather-robust and domain-invariant representations. Empirical results on SemanticKITTI, SynLiDAR, and SemanticSTF demonstrate state-of-the-art gains in both UDA and DG, with substantial improvements in mean IoU across diverse weather conditions and unseen scenes. Beyond adaptation, UniMix also serves as effective data augmentation for supervised LSS and shows potential for broader robustness in real-world autonomous driving deployments.

Abstract

LiDAR semantic segmentation (LSS) is a critical task in autonomous driving and has achieved promising progress. However, prior LSS methods are conventionally investigated and evaluated on datasets within the same domain in clear weather. The robustness of LSS models in unseen scenes and all weather conditions is crucial for ensuring safety and reliability in real applications. To this end, we propose UniMix, a universal method that enhances the adaptability and generalizability of LSS models. UniMix first leverages physically valid adverse weather simulation to construct a Bridge Domain, which serves to bridge the domain gap between the clear weather scenes and the adverse weather scenes. Then, a Universal Mixing operator is defined regarding spatial, intensity, and semantic distributions to create the intermediate domain with mixed samples from given domains. Integrating the proposed two techniques into a teacher-student framework, UniMix efficiently mitigates the domain gap and enables LSS models to learn weather-robust and domain-invariant representations. We devote UniMix to two main setups: 1) unsupervised domain adaption, adapting the model from the clear weather source domain to the adverse weather target domain; 2) domain generalization, learning a model that generalizes well to unseen scenes in adverse weather. Extensive experiments validate the effectiveness of UniMix across different tasks and datasets, all achieving superior performance over state-of-the-art methods. The code will be released.

Figures (4)

• Figure 1: UniMix outperforms SOTA methods in both UDA and DG tasks, using SemanticKITTI behley2019semantickitti as the source and SemanticSTF xiao20233d in all four weather conditions as the target.
• Figure 2: The overall pipeline of UniMix. In the first stage (top part), the clear-weather source domain $S_c$ and the simulated adverse-weather Bridge Domain $B_a$ are taken as input to generate the intermediate domain $\{P^{S_c \rightarrow B_a}, P^{B_a \rightarrow S_c}\}$ via Universal Mixing, and the student network is trained under the supervision of $L_{stage1}$. In the second stage (bottom part), the Bridge Domain and adverse-weather target domain $T_a$ are utilized to generate the intermediate domain $\{P^{B_a \rightarrow T_a}, P^{T_a \rightarrow B_a}\}$ via Universal Mixing, and the student network is trained under the supervision of $L_{stage2}$. The teacher is leveraged to produce pseudo labels and is updated via EMA tarvainen2017mean of the student's weights.
• Figure 3: Visualization of point clouds in the source domain, Bridge Domain, and target domain. The Bridge and target domain are in light fog weather. Better viewed with zoom-in and in color.
• Figure 4: Illustration of different sample mixing methods, including spatial mixing, intensity mixing, and semantic mixing.