ProxyMix: Proxy-based Mixup training with label refinery for source-free domain adaptation.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Source-free Unsupervised Domain Adaptation (SFDA) is crucial for transferring knowledge between domains without raw source data, addressing privacy and transmission concerns.
• Existing SFDA methods often require additional parameters or rely on noisy pseudo-labels, limiting their effectiveness.
• The need for efficient and accurate SFDA techniques is growing in various applications.

Purpose of the Study:

• To propose ProxyMix, an effective SFDA method that avoids additional parameters and noisy pseudo-labels.
• To leverage pre-trained source models and enhance knowledge transfer to unlabeled target domains.
• To improve the reliability of pseudo-labels and mitigate negative impacts of noisy data.

Main Methods:

• ProxyMix defines classifier weights as class prototypes and constructs a balanced proxy source domain using nearest neighbors.
• A frequency-weighted aggregation strategy generates reliable soft pseudo-labels for unlabeled target data.
• Inter- and intra-domain mixup regularization aligns proxy and target domains, utilizing target features' internal structure.

Main Results:

• ProxyMix achieves state-of-the-art performance on 2D image and 3D point cloud object recognition benchmarks.
• The proposed method effectively mitigates the negative impact of noisy labels through label refinery.
• ProxyMix demonstrates superior knowledge transfer capabilities in source-free domain adaptation scenarios.

Conclusions:

• ProxyMix offers an effective solution for Source-free Unsupervised Domain Adaptation, outperforming existing methods.
• The novel approach addresses key limitations of previous SFDA techniques, enhancing model robustness and accuracy.
• ProxyMix shows significant promise for real-world applications requiring privacy-preserving domain adaptation.