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

TL;DR

The coarse labels are introduced and the newly created multisource dataset COLA is called to overcome the common obstacles in multisource training and it is demonstrated that multisource approaches bring systematic improvement over single-source approaches.

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.

Figures (8)

• Figure 1: COLA and its applications: domain generalization (COLA-DG), source-to-source segmentation (COLA-S2S), and pre-training (COLA-PT). COLA-DG uses the mixed-domain training set to improve robustness and performance over unseen domains. COLA-S2S leverages the mixed-domain training set as a basis and complements it with a sample of the target set. COLA-PT uses the mixed-domain training set to extract a pre-trained model that can be finetuned on any target set.
• Figure 2: Illustration of naive multi-source methods.
• Figure 3: SemanticKITTI's label set and mapping to the coarse labels.
• Figure 4: Coarse label vehicle and the training labels that constitute it.
• Figure 5: Illustration of COLA-DG. It takes several datasets extracted from various domains as input and remaps them to a single label set through COLA. This new dataset is used to learn a domain-robust semantic segmentation model.