PointSFDA: Source-free Domain Adaptation for Point Cloud Completion

2025-03-19

AI Summary

Point cloud completion models trained on synthetic data suffer severe performance degradation on real-world scans due to domain shift. To address this, we propose the first source-free domain adaptation (SFDA) framework specifically designed for cross-domain point cloud completion—requiring no access to source-domain data. Our method integrates coarse-to-fine knowledge distillation with self-supervised local masked consistency learning: the former leverages a teacher model to guide hierarchical reconstruction by the student, while the latter enforces structural priors through unsupervised masked point cloud reconstruction. The resulting architecture is fully source-free, yet significantly improves both accuracy and robustness of state-of-the-art models on real-world scans. We achieve leading performance across multiple cross-domain benchmarks. Code is publicly available.

Technology Category

• Machine Learning: Transfer, Domain Adaptation, Multi-Task Learning
• Computer Vision: Diffusion Models for Vision
• Natural Language Processing: Code Generation / Program Synthesis from Natural Language

Application Category

• Search and Retrieval-Augmented AI: Vertical and domain-specific search
• Graph Algorithms and Modeling for the Web: Algorithms and analysis for incomplete, noisy, or partially observed Web-related graphs
• Web Mining and Content Analysis: Large pretrained models with web data

Abstract

Conventional methods for point cloud completion, typically trained on synthetic datasets, face significant challenges when applied to out-of-distribution real-world scans. In this paper, we propose an effective yet simple source-free domain adaptation framework for point cloud completion, termed extbf{PointSFDA}. Unlike unsupervised domain adaptation that reduces the domain gap by directly leveraging labeled source data, PointSFDA uses only a pretrained source model and unlabeled target data for adaptation, avoiding the need for inaccessible source data in practical scenarios. Being the first source-free domain adaptation architecture for point cloud completion, our method offers two core contributions. First, we introduce a coarse-to-fine distillation solution to explicitly transfer the global geometry knowledge learned from the source dataset. Second, as noise may be introduced due to domain gaps, we propose a self-supervised partial-mask consistency training strategy to learn local geometry information in the target domain. Extensive experiments have validated that our method significantly improves the performance of state-of-the-art networks in cross-domain shape completion. Our code is available at emph{ extcolor{magenta}{https://github.com/Starak-x/PointSFDA}}.

Problem

• Addresses domain gap in point cloud completion
• Uses pretrained source model and unlabeled target data
• Improves cross-domain shape completion performance

Innovation

• Source-free domain adaptation for point clouds
• Coarse-to-fine distillation for geometry transfer
• Self-supervised partial-mask consistency training