Abstract
Understanding point clouds captured from the real-world is challenging due to shifts in data distribution caused by varying object scales, sensor angles, and self-occlusion. Prior works have addressed this issue by combining recent learning principles such as self-supervised learning, self-training, and adversarial training, which leads to significant computational overhead. Toward succinct yet powerful domain adaptation for point clouds, we revisit the unique challenges of point cloud data under domain shift scenarios and discover the importance of the global geometry of source data and trends of target pseudo-labels biased to the source label distribution. Motivated by our observations, we propose an adapter-guided domain adaptation method, PC-Adapter, that preserves the global shape information of the source domain using an attention-based adapter, while learning the local characteristics of the target domain via another adapter equipped with graph convolution. Additionally, we propose a novel pseudo-labeling strategy resilient to the classifier bias by adjusting confidence scores using their class-wise confidence distributions to consider relative confidences. Our method demonstrates superiority over baselines on various domain shift settings in benchmark datasets - PointDA, GraspNetPC, and PointSegDA.
Introduction
In 3D computer vision, models trained on synthetic data (like clean CAD models) often fail when deployed in the real world using noisy LiDAR or depth scans. This "domain shift" is caused by differences in sensor noise, object scales, and occlusions. While Unsupervised Domain Adaptation (UDA) aims to bridge this gap, current methods are often computationally heavy, requiring massive resources for adversarial training or complex self-supervised tasks.
Furthermore, many UDA methods rely on "pseudo-labeling," which often fails because the model becomes biased toward the most common classes in the original training set. PC-Adapter addresses these issues through a parameter-efficient adapter approach and a novel label rectification strategy.
Methodology
The authors introduce PC-Adapter, which moves away from heavy retraining toward a dual-adapter strategy.
Dual-Adapter Architecture
Instead of modifying the entire backbone, the framework inserts two small, specialized modules:
- Global Shape-Aware Adapter ($\Psi_g$): Uses self-attention and a novel Relative Positional Encoding to capture the "skeleton" or general anatomy of an object. This module is trained to preserve robust geometric knowledge from the source domain.
- Locality-Aware Adapter ($\Psi_l$): Uses Graph Convolutions (GCN) to focus on the fine-grained, local structures unique to the target domain, such as specific sensor noise patterns.
Rectified Pseudo-labeling
To fix label bias, the authors model the model's confidence scores as a Beta distribution. Instead of just picking labels with the highest raw score, they adjust scores based on their "percentile rank" within each class. This allows the model to identify high-confidence examples for minority classes that would otherwise be ignored due to classifier overconfidence in major classes.
Key Insights and Ideas
A critical discovery of this research is that global shape is the most "transferable" type of knowledge. The authors found that even if 75% of the points in a cloud are removed (keeping only the global structure via Farthest Point Sampling), classification accuracy under domain shift often remains stable or even improves.
By using distribution-guided rectification, PC-Adapter can generate unbiased pseudo-labels without needing to know the target domain's true label distribution in advance. This efficiency makes the method significantly faster and more practical for real-world systems with limited compute.
Key Results
The researchers evaluated PC-Adapter across three major benchmarks: PointDA-10, GraspNetPC-10, and PointSegDA.
- Superior Accuracy: PC-Adapter achieved an average accuracy of 70.3% on PointDA-10, significantly outperforming the previous state-of-the-art (GLRV at 62.7%).
- Segmentation Boost: On the PointSegDA task, it showed a 3.5% mIoU (Mean Intersection over Union) enhancement over baselines.
- Speedup: The training process is roughly 3.43x faster than leading self-supervised learning methods.
- Data Scarcity: The model remains robust even when the amount of available target data is slashed to just 10%, whereas baseline performance drops sharply.
Conclusion
PC-Adapter provides a lightweight, high-performance framework for 3D point cloud domain adaptation. By balancing global shape preservation with local feature learning and rectifying pseudo-label bias, it achieves state-of-the-art results with minimal computational overhead. Future work includes extending this approach to complex 3D scenes and real-time robotics applications.
Summary updated