Self-Supervised LiDAR Desnowing with 3D-KNN Blind-Spot Networks

Light Detection and Ranging (LiDAR) is fundamental to autonomous driving and robotics, as it provides reliable 3D geometric information. However, snowfall introduces numerous spurious reflections that corrupt range measurements and severely degrade downstream perception. Existing desnowing techniques either rely on handcrafted filtering rules that fail under varying snow densities, or require paired snowy–clean scans, which are nearly impossible to collect in real-world scenarios. Self-supervised LiDAR desnowing approaches address these challenges by projecting raw 3D point clouds into 2D range images and jointly training a point reconstruction network (PR-Net) and a reconstruction difficulty network (RD-Net). Nevertheless, these methods remain limited by their reliance on the outdated Noise2Void training paradigm, which restricts reconstruction quality. In this paper, we redesign PR-Net with a blind-spot architecture to overcome the limitation. Specifically, we introduce a 3D-KNN encoder that aggregates neighborhood features directly in Euclidean 3D space, ensuring geometrically consistent representations. Additionally, we integrate residual state-space blocks (RSSB) to capture long-range contextual dependencies with linear computational complexity. Extensive experiments on both synthetic and real-world datasets, including SnowyKITTI and WADS, demonstrate that our method outperforms state-of-the-art self-supervised desnowing approaches by up to 0.06 IoU while maintaining high computational efficiency.