Enhancing LiDAR Point Cloud Semantic Segmentation via Multiscale Multiview Fusion and Knowledge Distillation

LiDAR point cloud semantic segmentation is essential for tasks such as autonomous driving and urban modeling. However, existing multimodal fusion methods often suffer from information loss due to alignment issues and field-of-view discrepancies. To address these challenges, we propose a novel multiscale multiview fusion network, MFKDSeg, leveraging only single-modal LiDAR point clouds. Our approach integrates voxel index-based feature aggregation (VIFA) and dual-channel point-to-voxel fusion-to-single voxel distillation (DPVtD). VIFA efficiently aggregates point-wise features at different scales using voxel indices, while DPVtD incorporates attention-based fusion (AF) and uncertainty-based fusion (UF) mechanisms to provide multilevel knowledge to the voxel network. Here, we show that MFKDSeg achieves state-of-the-art (SOTA) performance on two large benchmark datasets, SemanticKITTI and nuScenes, demonstrating its effectiveness in enhancing LiDAR semantic segmentation for large-scale scenes.