NormFusion: Real-Time Test-Time Adaptation for LiDAR Segmentation under Corruption

Real-time adaptation of LiDAR-based perception models is essential for safety-critical applications such as autonomous driving, where models must remain robust in diverse and dynamic environments. Current test-time adaptation approaches largely overlook point cloud data and struggle to meet the real-time demands of frame-by-frame updates. To address these limitations, we introduce Real-Time Test-Time Adaptation (RTTA) and propose a novel method called NormFusion. NormFusion enhances traditional batch normalization by incorporating the fusion of normalization statistics, effectively handling distribution shifts encountered during real-time operation. Unlike directly interpolation, our method applies distribution fusion according to network stages, ensuring stable and efficient adaptation without adding computational costs overhead. Experimental results on LiDAR semantic segmentation tasks, including the nuScenes and SemanticKITTI datasets, demonstrate that NormFusion outperforms state-of-the-art methods in both accuracy and computational efficiency, significantly enhancing performance under corrupted, real-world conditions.