Toward Enhanced Representation Learning for Single-Source Domain Generalization in LiDAR Semantic Segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• 3D deep learning models excel in LiDAR semantic segmentation but struggle with domain shifts in unseen environments.
• Domain generalization is crucial for robust autonomous driving perception systems.
• Existing methods often fail to generalize effectively from a single source domain to diverse, real-world scenarios.

Purpose of the Study:

• To propose DGLSS++, a representation learning approach for domain generalization in LiDAR semantic segmentation.
• To ensure robust performance in both source and unseen domains, even when trained solely on the source domain.
• To address domain shifts caused by variations in LiDAR sensor configurations and scene distributions.

Main Methods:

• Developed DGLSS++ for single-source domain generalization in LiDAR semantic segmentation.
• Simulated unseen domains (sparse-to-dense and dense-to-sparse) via augmentation.
• Introduced generalized masked sparsity invariant feature consistency (GMSIFC) and localized semantic correlation consistency (LSCC) for generalizable representation learning.
• GMSIFC aligns sparse features across domains with a novel masking strategy; LSCC maintains local semantic correlations.

Main Results:

• DGLSS++ demonstrated superior performance compared to Unsupervised Domain Adaptation (UDA) and Domain Generalization (DG) baselines.
• The approach achieved robust performance across various LiDAR sensor configurations and scene distributions.
• Experiments utilized four real-world datasets, validating the effectiveness of the proposed methods.

Conclusions:

• DGLSS++ effectively addresses the domain gap in LiDAR semantic segmentation for autonomous driving.
• The proposed constraints (GMSIFC and LSCC) facilitate robust generalization from a single source domain.
• The developed approach offers a promising solution for real-world autonomous driving perception systems.