PointSFDA: Source-Free Domain Adaptation for Point Cloud Completion
Xing He, Zhe Zhu, Liangliang Nan, Wenshuo Peng, Honghua Chen, Mingqiang Wei
AI summary
Problem
Deep learning models for point cloud completion degrade under domain gaps between synthetic training data and real-world scans, while existing unsupervised domain adaptation methods require inaccessible source data.
Approach
PointSFDA adapts a pre-trained model using only unlabeled target data by distilling global geometric priors from the source model and enforcing local geometric consistency through masked self-supervised training.
Key results
- First source-free domain adaptation framework for point cloud completion
- Coarse-to-fine point cloud distillation transfers domain-invariant global geometry
- Partial-mask consistency training enables self-supervised local geometry learning
- Achieves significant Chamfer distance reductions on KITTI, ScanNet, 3D-FUTURE, and ModelNet datasets
Why it matters
Provides a practical, scalable solution for autonomous driving and robotics applications where source training data is proprietary or inaccessible.
Abstract
Point cloud completion is critical for autonomous driving and robotic perception, yet deep learning models often experience severe performance degradation under the domain gap between synthetic training and real-world data. While unsupervised domain adaptation (UDA) has been explored to mitigate this issue, its reliance on access to source datasets limits practical applicability, as source data are often proprietary or restricted. We pioneer source-free domain adaptation (SFDA) for point cloud completion, which adapts a pre-trained source model to an unlabeled target domain without requiring source data access. To this end, we propose PointSFDA, a framework that combines global knowledge transfer with target-specific lo- cal adaptation. Specifically, we design (i) a Coarse-to-Fine Point Cloud Distillation module to extract domain-invariant global geometric priors from the source model, and (ii) a Partial-Mask Consistency Training strategy to enforce prediction consistency across masking augmentations, enabling self-supervised learn- ing of local target-domain geometry. Experiments on real-world datasets (KITTI, ScanNet) and synthetic benchmarks (Mod- elNet40, 3D-FUTURE) demonstrate that PointSFDA achieves significant improvements over state-of-the-art methods in cross- domain shape completion, establishing a practical and scalable solution for robotics applications. Our code is available at https://github.com/Starak-x/PointSFDA.