PhyDAWS: Physically-inspired data augmentation with weather simulation for domain-generalized point cloud segmentation

Point cloud semantic segmentation performance in autonomous driving systems degrades notably under adverse weather conditions, posing a challenge for reliable deployment. Current data augmentation methods for domain generalization often inadequately model the physical interactions between LiDAR sensors and atmospheric particles specific to diverse weather phenomena. This paper introduces PhyDAWS, a physically-inspired data augmentation framework designed for domain generalization in point cloud segmentation. It aims to enhance model robustness across various weather conditions using only source domain data for training. PhyDAWS incorporates two complementary weather simulation techniques. The first is a phenomenological method modeling angle-dependent occlusion (rain), height-dependent density variation (snow), and distance-dependent attenuation (fog). The second is a simulation based on Mie scattering theory, accounting for particle size distributions, refractive indices, and wavelength-dependent scattering. These physically-inspired augmentations are integrated with a dual-view contrastive learning strategy to promote the extraction of weather-invariant features. Evaluations on domain generalization benchmarks, including SemanticKITTI to SemanticSTF and SynLiDAR to SemanticSTF transfers, demonstrate the framework’s effectiveness. Specifically for the SemanticKITTI to SemanticSTF task, PhyDAWS achieves improvements in overall mean Intersection over Union (mIoU) ranging from 4.1% to 10.3% compared to recent state-of-the-art approaches. By bridging the domain gap between clear and adverse weather conditions, this method advances the capability of autonomous systems for reliable environmental perception under diverse, unseen adverse conditions. • Physically-inspired weather simulation for domain-generalized LiDAR segmentation. • Phenomenological and Mie scattering models for realistic point cloud augmentation. • Robust perception framework across diverse and unseen adverse weather conditions.