LEST: Large-scale LiDAR Semantic Segmentation with Transformer

TL;DR

This paper introduces LEST, a novel Transformer-based architecture for large-scale LiDAR point cloud semantic segmentation, outperforming existing methods on major benchmarks.

Contribution

The paper presents a new Transformer architecture with SFC grouping and DISCO components for improved LiDAR segmentation.

Findings

Outperforms state-of-the-art methods on nuScenes and SemanticKITTI datasets

Introduces SFC grouping strategy for efficient point cloud processing

Develops Distance-based Cosine Linear Transformer (DISCO) for better feature extraction

Abstract

Large-scale LiDAR-based point cloud semantic segmentation is a critical task in autonomous driving perception. Almost all of the previous state-of-the-art LiDAR semantic segmentation methods are variants of sparse 3D convolution. Although the Transformer architecture is becoming popular in the field of natural language processing and 2D computer vision, its application to large-scale point cloud semantic segmentation is still limited. In this paper, we propose a LiDAR sEmantic Segmentation architecture with pure Transformer, LEST. LEST comprises two novel components: a Space Filling Curve (SFC) Grouping strategy and a Distance-based Cosine Linear Transformer, DISCO. On the public nuScenes semantic segmentation validation set and SemanticKITTI test set, our model outperforms all the other state-of-the-art methods.