SnowRemovalNet: BEV-Guided 3D Point Feature Learning for LiDAR De-Snowing in Autonomous Driving

In heavy snowfall, LiDAR point clouds contain numerous noise points caused by scattering and reflection from snow particles. In particular, snow and non-snow points are often spatially intermingled, making fine-grained pointwise discrimination challenging. However, existing LiDAR de-snowing methods have not sufficiently focused on this problem. Therefore, this study proposes SnowRemovalNet, a supervised LiDAR de-snowing network designed for fine-grained pointwise discrimination in complex snowy scenes. The proposed method incorporates global contextual information extracted from a bird's-eye view (BEV) representation into 3D point feature learning. This enables the network to jointly consider the spatial distribution of snow noise and the local geometric relationships of individual points. To account for label uncertainty near the boundaries between snow and non-snow regions, the snow-noise likelihood of each point is represented as a continuous probability between 0 and 1. This value serves as the probabilistic label during training. On the CADC and WADS datasets, the proposed model effectively distinguishes spatially intermingled snow and non-snow points. Using only 30 training frames per dataset, the model achieves F1-scores of 98.10% on CADC and 97.23% on WADS. The processing time is 17-59 ms per frame.