Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

DPGLA: Bridging the Gap between Synthetic and Real Data for Unsupervised Domain Adaptation in 3D LiDAR Semantic Segmentation

Wanmeng Li, Simone Mosco, Daniel Fusaro, Alberto Pretto

TL;DR

DPGLA tackles unsupervised domain adaptation for 3D LiDAR semantic segmentation by integrating a dynamic pseudo-label filtering scheme (DPLF), a non-learned prior-guided data augmentation pipeline (PG-DAP), and a data mixing consistency loss within a Mean Teacher framework. DPLF adapts global and class-specific confidence thresholds using EMA-driven statistics to balance pseudo-label quality and class representation, while PG-DAP employs Density-Aware Sampling and jitter strategies to bridge input-level disparities without heavy computation. The data mixing consistency loss enforces context-free representations on mixed source/target samples, and LaserMix-based domain mixing further promotes robust cross-domain learning. Across two synthetic-to-real benchmarks, DPGLA achieves state-of-the-art performance and ablations show each component contributes to improved robustness and accuracy, with released code enabling reproducibility and future extensions.

Abstract

Annotating real-world LiDAR point clouds for use in intelligent autonomous systems is costly. To overcome this limitation, self-training-based Unsupervised Domain Adaptation (UDA) has been widely used to improve point cloud semantic segmentation by leveraging synthetic point cloud data. However, we argue that existing methods do not effectively utilize unlabeled data, as they either rely on predefined or fixed confidence thresholds, resulting in suboptimal performance. In this paper, we propose a Dynamic Pseudo-Label Filtering (DPLF) scheme to enhance real data utilization in point cloud UDA semantic segmentation. Additionally, we design a simple and efficient Prior-Guided Data Augmentation Pipeline (PG-DAP) to mitigate domain shift between synthetic and real-world point clouds. Finally, we utilize data mixing consistency loss to push the model to learn context-free representations. We implement and thoroughly evaluate our approach through extensive comparisons with state-of-the-art methods. Experiments on two challenging synthetic-to-real point cloud semantic segmentation tasks demonstrate that our approach achieves superior performance. Ablation studies confirm the effectiveness of the DPLF and PG-DAP modules. We release the code of our method in this paper.

DPGLA: Bridging the Gap between Synthetic and Real Data for Unsupervised Domain Adaptation in 3D LiDAR Semantic Segmentation

TL;DR

DPGLA tackles unsupervised domain adaptation for 3D LiDAR semantic segmentation by integrating a dynamic pseudo-label filtering scheme (DPLF), a non-learned prior-guided data augmentation pipeline (PG-DAP), and a data mixing consistency loss within a Mean Teacher framework. DPLF adapts global and class-specific confidence thresholds using EMA-driven statistics to balance pseudo-label quality and class representation, while PG-DAP employs Density-Aware Sampling and jitter strategies to bridge input-level disparities without heavy computation. The data mixing consistency loss enforces context-free representations on mixed source/target samples, and LaserMix-based domain mixing further promotes robust cross-domain learning. Across two synthetic-to-real benchmarks, DPGLA achieves state-of-the-art performance and ablations show each component contributes to improved robustness and accuracy, with released code enabling reproducibility and future extensions.

Abstract

Annotating real-world LiDAR point clouds for use in intelligent autonomous systems is costly. To overcome this limitation, self-training-based Unsupervised Domain Adaptation (UDA) has been widely used to improve point cloud semantic segmentation by leveraging synthetic point cloud data. However, we argue that existing methods do not effectively utilize unlabeled data, as they either rely on predefined or fixed confidence thresholds, resulting in suboptimal performance. In this paper, we propose a Dynamic Pseudo-Label Filtering (DPLF) scheme to enhance real data utilization in point cloud UDA semantic segmentation. Additionally, we design a simple and efficient Prior-Guided Data Augmentation Pipeline (PG-DAP) to mitigate domain shift between synthetic and real-world point clouds. Finally, we utilize data mixing consistency loss to push the model to learn context-free representations. We implement and thoroughly evaluate our approach through extensive comparisons with state-of-the-art methods. Experiments on two challenging synthetic-to-real point cloud semantic segmentation tasks demonstrate that our approach achieves superior performance. Ablation studies confirm the effectiveness of the DPLF and PG-DAP modules. We release the code of our method in this paper.

Paper Structure

This paper contains 15 sections, 23 equations, 2 figures, 3 tables.

Table of Contents

  1. INTRODUCTION
  2. RELATED WORK
  3. Our Approach
  4. Preliminaries
  5. Architecture
  6. Dynamic Pseudo-Label Filtering
  7. Prior-Guided Data Augmentation Pipeline
  8. Data Mixing Consistency
  9. Network Update
  10. EXPERIMENTAL EVALUATION
  11. Datasets and Evaluation Metrics
  12. Implementation Details
  13. Point Cloud UDA Semantic Segmentation
  14. Ablation Studies
  15. CONCLUSIONS

Figures (2)

  • Figure 1: Confidence distribution of pseudo-labels for different classes on sequence 01 of the SemanticKITTI 9010727. The gray area represents real point cloud data excluded from subsequent training, while the white area indicates the retained data. The red line represents the fixed confidence threshold in CoSMix. The green and blue lines denote global and class-specific thresholds of DPLF in DPGLA, respectively.
  • Figure 2: Overview of DPGLA architecture. The model takes as input a pair of sample point clouds: one labeled from source domain $\mathcal{D}_s$ (upper branch) and one unlabeled from target domain $\mathcal{D}_t$ (bottom branch).