Adverse weather (rain, fog, snow) corrupts LiDAR through backscatter and attenuation, producing near-range clutter and far-range sparsity that degrade downstream perception and planning. Classical geometric filters depend on hand-tuned thresholds and often erode valid structure at long range; recent learning methods improve accuracy but are heavy for embedded deployment. We present LRINet, a straightforward and efficient range-image network for denoising LiDAR point clouds in adverse weather. LRINet operates entirely in a 2-D cylindrical projection. This approach preserves sensor topology and ensures computational efficiency. The network pairs depth-guided attention with a coordinate encoder and a global-context block. It then uses an attention-gated decoder for the final output. Despite its minimalist design, LRINet delivers competitive performance against state-of-the-art methods across rainy, foggy, and snowy conditions while maintaining a markedly smaller memory and latency footprint suitable for real-time on-board use. Experiments on climate-chamber rain/fog data and semi-synthetic snowy driving data show a practical balance of accuracy, robustness, and throughput, making LRINet a drop-in preprocessing module for safety-critical autonomy stacks.
La et al. (Thu,) studied this question.