Source-Only Cross-Weather LiDAR via Geometry-Aware Point Drop

TL;DR

The paper tackles LiDAR semantic segmentation under adverse weather by addressing boundary and sparse-region vulnerabilities through a Light Geometry-aware adapter. This module computes geometry cues via a local-window KNN with horizontal circular padding and integrates them into a region-aware regularization framework that guides region-level dropping before Learnable Point Drop. In a source-only cross-weather setup (SemanticKITTI→SemanticSTF), the approach yields substantial gains over data-centric augmentation and modest gains over class-centric regularization, validating geometry-driven regularization as a key direction for robust all-weather LiDAR segmentation. Its plug-and-play nature and negligible inference cost support practical deployment in autonomous systems facing weather-induced disturbances.

Abstract

LiDAR semantic segmentation degrades in adverse weather because refraction, scattering, and point dropouts corrupt geometry. Prior work in weather simulation, mixing-based augmentation, domain randomization, and uncertainty or boundary regularization improves robustness but still overlooks structural vulnerabilities near boundaries, corners, and sparse regions. We present a Light Geometry-aware adapter. The module aligns azimuth and applies horizontal circular padding to preserve neighbor continuity across the 0~360 degree wrap-around boundary. A local-window K-Nearest Neighbors gathers nearby points and computes simple local statistics, which are compressed into compact geometry-aware cues. During training, these cues drive region-aware regularization that stabilizes predictions in structurally fragile areas. The adapter is plug and play, complements augmentation, and can be enabled only during training with negligible inference cost. We adopt a source-only cross-weather setup where models train on SemanticKITTI and are evaluated on SemanticSTF without target labels or fine-tuning. The adapter improves mIoU by 7.9 percentage points over the data-centric augmentation baseline and by 0.6 points over the class-centric regularization baseline. These results indicate that geometry-driven regularization is a key direction for all-weather LiDAR segmentation.

Figures (5)

• Figure 1: Data-centric augmentation baseline in adverse weather. Left (a) the baseline merges a bus into a nearby building, while our Light Geometry-aware adapter preserves neighbor continuity and segments the bus (blue) instead of the building (yellow). Right (b) the baseline blurs the building–fence boundary, while ours keeps the boundary and labels building (yellow) instead of fence (orange).
• Figure 2: Three LiDAR segmentation pipelines under the source-only training setting. (a) Conventional networks without noise or structure modeling. (b) Data-centric augmentation with selective jittering and a learnable point drop module. (c) Ours with a Light Geometry-aware adapter, which enhances region-level dropping by preserving local geometry.
• Figure 3: Overall pipeline. Selective jittering perturbs inputs to emulate weather-induced corruptions. A Light Geometry-aware adapter, inserted before learnable point drop, derives compact geometry-aware cues via local-window KNN with circular wrapping and preserves $0^\circ$--$360^\circ$ continuity. These cues guide the drop policy toward region-level dropping at boundaries, corners, and sparse structures. After dropping, the backbone processes the perturbed input and produces semantic predictions with improved robustness under adverse weather.
• Figure 4: Global KNN builds neighbors from the full scan and often pulls distant points near the azimuth seam, which increases computation and weakens locality. Our Light Geometry-aware adapter uses a horizontal local-window KNN with circular padding to wrap the $0^\circ$–$360^\circ$ seam, preserve edge and corner cues, and reduce computation. This design makes the module light.
• Figure 5: Qualitative comparison on SemanticSTF across four adverse-weather conditions: light fog, dense fog, rain, and snow. The top row shows the data-centric augmentation baseline. The middle and bottom rows show our method with $(W=512, K=8)$ and $(W=256, K=16)$. The Light Geometry-aware adapter preserves neighbor continuity and reduces class confusion at boundaries.