A2CFlow: Range-Aware Rectified Flow for Efficient Adverse-Weather LiDAR Restoration
Published in Revised and resubmitted to IEEE Robotics and Automation Letters (RA-L), 2026
Abstract
Adverse weather degrades LiDAR measurements nonuniformly, leaving distant regions and missing-return locations weakly constrained by the corrupted observation. Effective restoration must therefore suppress weather-induced clutter while reconstructing clean geometry that is poorly observed or absent. We introduce Adverse-to-Clear Flow (A2CFlow), a conditional rectified-flow framework that generates a clean range-and-reflectance scan from Gaussian noise conditioned on a single adverse-weather observation. To account for spatially varying restoration difficulty, we propose Range-Aware Flow Scheduling (RAFS), a physics-informed reparameterization derived from observed range and return validity. RAFS assigns location-dependent flow times, providing difficult regions with cleaner training states and greater early sampling progress without changing the transport endpoints or increasing network evaluations. We further compress the transport into an efficient few-step distilled model that retains the benefits of range-aware scheduling. Extensive experiments demonstrate improved restoration fidelity and downstream perception on Adverse-KITTI, as well as zero-shot transfer to real adverse-weather scans, while retaining efficient inference.
Recommended citation: Peiyang Ni, Xiaorui Sun, Lu Zhang, Heng Tao Shen, Xiaofeng Zhu, and Ping Hu. "A2CFlow: Range-Aware Rectified Flow for Efficient Adverse-Weather LiDAR Restoration." Revised and resubmitted to IEEE Robotics and Automation Letters (RA-L), 2026.