PointSFDA: Source-free Domain Adaptation for Point Cloud Completion

TL;DR

PointSFDA tackles the domain gap in point cloud completion by performing source-free domain adaptation using only a pretrained source model and unlabeled target data. It introduces coarse-to-fine point cloud distillation to transfer global geometry and partial-mask consistency training to learn local target geometry, complemented by an EMA-based mutual refinement of the source model. The method shows substantial cross-domain improvements on real (KITTI, ScanNet) and synthetic (3D-FUTURE, ModelNet40) benchmarks, outperforming existing UDA and unsupervised methods. The work provides a practical, data-access-friendly approach with open-source code that can benefit cross-domain point cloud completion tasks.

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 \textbf{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{\textcolor{magenta}{https://github.com/Starak-x/PointSFDA}}.

Figures (6)

• Figure 1: Illustration of (a) unsupervised domain adaptation (UDA) and (b) source-free domain adaptation (SFDA) in point cloud completion. The UDA methods rely on both source and target datasets during adaptation, while SFDA methods only have access to a pretrained source model and unlabeled target data. Both UDA and SFDA aim to generate complete point clouds that resemble the distribution of real data.
• Figure 2: Overview of the proposed PointSFDA. A partial point cloud from the target domain is directly used as input for the source model and masked $k$ times ($k = 2$ in this figure) before being fed into the target model. Coarse-to-fine Point Cloud Distillation is employed in the coarse point cloud and fine point cloud to directly transfer the geometry information across domains. Mask-partial Consistency Training learns target data geometry information through point-wise predictive consistency across various mask augmentations.
• Figure 3: Visual comparison with recent methods(ACL-SPC ACLSPC, P2C P2C, OptDE OptDE, AdaPoinTr AdaPoinTr) on the test set of KITTI and ScanNet.
• Figure 4: Visual comparison with recent methods(ACL-SPC ACLSPC, P2C P2C, OptDE OptDE, SnowflakeNet SnowflakeNet) on the test set of 3D-FUTURE and ModelNet. The plane and car come from ModelNet, while the chair, sofa, and table are from 3D-FUTURE.
• Figure 5: The ablation study of PMCT. In Variant B, only coarse-to-fine point cloud distillation is used, without employing PMCT. Notably, Variant B may recover false chair legs due to the domain gap.