AI summary
Problem
Adverse weather severely degrades LiDAR semantic segmentation by causing point dropouts and distorting object boundaries, yet existing augmentation methods often ignore these structural vulnerabilities.
Approach
The authors introduce a plug-and-play adapter that preserves LiDAR neighbor continuity and extracts local geometric cues to guide reinforcement learning-based region dropping during training.
Key results
- +3.4 mIoU gain over strong data-centric baselines on SemanticKITTI→SemanticSTF
- Performance matches advanced class-centric regularization methods
- Plug-and-play design adds negligible inference overhead
- Effectively suppresses boundary confusion and label inversions in fog, rain, and snow
Why it matters
Enables robust, all-weather LiDAR perception for autonomous driving using only source-domain training data, eliminating the need for costly target-domain collection or fine-tuning.
Abstract
Adverse weather conditions, such as rain, snow, and fog, severely degrade LiDAR semantic segmentation by introducing refraction, scattering, and point dropouts that com- promise geometric integrity. While prior approaches ranging from weather simulation and mixing-based augmentation to domain randomization and regularization enhance robustness, they frequently overlook structural vulnerabilities inherent to object boundaries, corners, and highly sparse regions. To address this limitation, we propose a Light Geometry- Aware Adapter. This module aligns azimuths and applies hori- zontal circular padding to preserve neighbor continuity across the 0◦–360◦wrap-around boundary. Using a local-window K- Nearest Neighbors (KNN) search, it aggregates nearby points and computes lightweight local statistics, compressing them into compact geometry-aware cues. During training, these cues facilitate region-aware regularization, which effectively stabilizes predictions in structurally fragile areas. The proposed adapter is designed to be plug-and-play, complements existing augmentation techniques, and operates exclusively during train- ing, incurring negligible inference overhead. Operating under a rigorous source-only cross-weather paradigm wherein models are trained on SemanticKITTI and evaluated on SemanticSTF without target-domain labels or fine-tuning, our adapter achieves a +3.4 mIoU improvement over strong data-centric augmentation baselines. Furthermore, it demonstrates performance comparable to advanced class- centric regularization methods. These findings highlight that geometry-driven regularization constitutes a critical pathway toward achieving highly robust, all-weather LiDAR segmenta- tion.