Enhancing LiDAR Point Cloud Semantic Segmentation via Multi-scale Multi-view Fusion and Knowledge Distillation

LiDAR point cloud semantic segmentation is essential for tasks such as autonomous driving and urban modeling. However, existing multi-modal fusion methods often suffer from information loss due to alignment issues and field-of-view discrepancies. To address these challenges, we propose a novel multi-scale multi-view fusion network, MV3Dseg, leveraging only single-modal LiDAR point clouds. Our approach integrates Voxel-based Indexing Feature Aggregation (VIFA) and Dual-channel Point-Voxel Fusion-to-Single Voxel Distillation (DPVtD). VIFA efficiently aggregates point-wise features at different scales using voxel indices, while DPVtD incorporates Attention-based and Uncertainty-based Fusion mechanisms to provide multi-level knowledge to the voxel network. Here, we show that MV3Dseg achieves state-of-the-art performance on two large benchmark datasets, SemanticKITTI and nuScenes, demonstrating its effectiveness in enhancing LiDAR semantic segmentation for large-scale scenes.