LEST: Large-Scale LiDAR Semantic Segmentation With Deployment-Friendly Transformer Architecture

Abstract

Large-scale LiDAR-based point cloud semantic segmentation is a critical challenge for autonomous driving perception. Most state-of-the-art LiDAR semantic segmentation methods rely on complex operators, such as sparse 3D convolutions or KdTree structures, which hinder their deployment on modern embedded devices. While transformer architectures have gained prominence in natural language processing (NLP) and 2D computer vision, their application to large-scale point cloud semantic segmentation remains limited. In this paper, we introduce LEST (LiDAR sEmantic Segmentation architecture with Transformer), a novel framework built entirely on simple operators. LEST incorporates two key innovations: a Space Filling Curve (SFC) grouping strategy and a DIStance-based COsine (DISCO) linear transformer. Experimental results demonstrate that our model achieves competitive performance on the nuScenes semantic segmentation validation set and the SemanticKITTI test set, while maintaining a deployment-friendly design.

Keywords

• Point cloud semantic segmentation
• representation learning
• long sequence modeling
• linear transformer
• model deployment