A Multi-View-Assisted Semantic Segmentation Network on LiDAR via Multi-Level Mutual Learning Knowledge Distillation

Abstract

LiDAR-based semantic segmentation is crucial in many robotic perception systems. Considering the data of LIDAR has various views with diverse spatial features, more and more semantic segmentation methods have been proposed to fuse them for better segmentation accuracy. However, compared to single-view segmentation, these multi-view methods that directly fuse all features face inferior real-time performance and higher computation costs. Therefore, this paper proposes a multi-view-assisted semantic segmentation network via multi-level mutual learning knowledge distillation (KD) to implement a high real-time and accurate semantic segmentation at a lower cost. The keys of the multi-level mutual learning-based KD strategy are intra-view mutual learning and inter-view mutual learning. To accelerate the computation and improve accuracy, we also introduce a two-step fusion strategy to fuse features hierarchically. Finally, we evaluated our approach on the SemanticKITTI dataset. The experimental results demonstrate that the proposed method competently improves efficiency and accuracy.