3DLabelProp: Geometric-Driven Domain Generalization for LiDAR Semantic Segmentation in Autonomous Driving

TL;DR

The paper tackles domain generalization for LiDAR semantic segmentation by introducing 3DLabelProp, a geometry-driven method that uses pseudo-dense point clouds generated from sequential scans to reduce sensor-domain discrepancies. It combines a fast, propagation-based geometry module for static elements with a learning-based cluster refinement (KPConv) for complex regions, and it fuses these predictions to deliver robust cross-dataset performance. A thorough DG benchmark across seven real datasets, along with ablations and comparisons to C&L, LiDOG, and DGLSS, shows that 3DLabelProp achieves strong generalization, particularly under sensor shifts, while maintaining competitive source-to-source performance. The work also discusses practical limitations (e.g., not real-time) and suggests directions for faster backbones and further optimization, highlighting its potential impact on robust autonomous driving perception.

Abstract

Domain generalization aims to find ways for deep learning models to maintain their performance despite significant domain shifts between training and inference datasets. This is particularly important for models that need to be robust or are costly to train. LiDAR perception in autonomous driving is impacted by both of these concerns, leading to the emergence of various approaches. This work addresses the challenge by proposing a geometry-based approach, leveraging the sequential structure of LiDAR sensors, which sets it apart from the learning-based methods commonly found in the literature. The proposed method, called 3DLabelProp, is applied on the task of LiDAR Semantic Segmentation (LSS). Through extensive experimentation on seven datasets, it is demonstrated to be a state-of-the-art approach, outperforming both naive and other domain generalization methods.

Figures (11)

• Figure 1: Illustration of our approach using pseudo-dense points for domain generalization of LiDAR semantic Segmentation in autonomous driving (the blue sphere represents the position of the ego vehicle).
• Figure 2: Illustration of appearance shift: On the left, an object labeled as a truck from the SemanticKITTI dataset semantickitti in Germany, and on the right, an object labeled as a truck from the nuScenes dataset nuscenes in the US.
• Figure 3: Illustration of scene shift: On the left, a scan from the SemanticKITTI dataset semantickitti in a German suburban area, and on the right, a scan from the SemanticPOSS dataset semanticposs on a university campus. Pedestrians are highlighted in green on the left, where they are located on sidewalks, and in blue on the right, where they are dispersed throughout the scene.
• Figure 4: Illustration of sensor shift: Both scans were acquired simultaneously from the PandaSet dataset pandaset. On the left is a scan from a solid-state LiDAR, and on the right is a scan from a 64-beam rotating LiDAR.
• Figure 5: Illustration of the trail phenomenon. On the left, a section of a point cloud made up of 5 consecutive scans; on the right, a section composed of 20 consecutive scans.