Dual stream knowledge distillation for 3D LiDAR segmentation model with large sparse multi-directional convolution

Semantic segmentation of LiDAR point clouds is a fundamental task in 3D scene understanding, with broad applications in autonomous driving, robotics, and urban mapping. This paper presents a novel framework that couples 3D large sparse multi-directional convolution with a training-time dual-stream knowledge distillation strategy. The backbone employs dynamic large sparse kernels and multi-directional convolution to create large effective receptive fields, while radially non-uniform cylindrical voxelization balances point density across the scene. During training, knowledge is distilled from two auxiliary modalities—2D RGB images and 1D space-filling-curve serialized point clouds—into the 3D backbone via a multi-layer bidirectional polarity-aware linear cross-attention mechanism, adding zero inference cost. Experiments on SemanticKITTI and nuScenes achieve mean IoU of 74.4% and 83.5%, respectively.