COLA: COarse-LAbel multi-source LiDAR semantic segmentation for autonomous driving

TL;DR

This paper introduces COLA, a multi-source LiDAR dataset with coarse labels, and demonstrates that multi-source training improves semantic segmentation performance across domain generalization, source-to-source segmentation, and pre-training tasks.

Contribution

The paper presents COLA, a novel multi-source LiDAR dataset with coarse labels, and shows systematic performance improvements in various segmentation tasks using multi-source training.

Findings

COLA improves domain generalization by +10%.

COLA enhances source-to-source segmentation by +5.3%.

COLA boosts pre-training results by +12%.

Abstract

LiDAR semantic segmentation for autonomous driving has been a growing field of interest in recent years. Datasets and methods have appeared and expanded very quickly, but methods have not been updated to exploit this new data availability and rely on the same classical datasets. Different ways of performing LIDAR semantic segmentation training and inference can be divided into several subfields, which include the following: domain generalization, source-to-source segmentation, and pre-training. In this work, we aim to improve results in all of these subfields with the novel approach of multi-source training. Multi-source training relies on the availability of various datasets at training time. To overcome the common obstacles in multi-source training, we introduce the coarse labels and call the newly created multi-source dataset COLA. We propose three applications of this new dataset that display systematic improvement over single-source strategies: COLA-DG for domain generalization (+10%), COLA-S2S for source-to-source segmentation (+5.3%), and COLA-PT for pre-training (+12%). We demonstrate that multi-source approaches bring systematic improvement over single-source approaches.