Feature Alignment by Uncertainty and Self-Training for Source-Free Unsupervised Domain Adaptation

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Unsupervised Domain Adaptation (UDA) typically requires labeled source data, which is often impractical due to privacy or storage limitations.
• Existing source-free UDA methods struggle with performance on perturbed images.

Purpose of the Study:

• To propose a novel source-free unsupervised domain adaptation method that overcomes limitations of existing approaches.
• To enhance model robustness against image perturbations in domain adaptation tasks.

Main Methods:

• The method utilizes only a pre-trained source model and unlabeled target images.
• It incorporates aleatoric uncertainty via data augmentation and trains the feature generator with dual consistency objectives.
• Leverages self-supervised learning principles for inter-space and intra-space alignment to minimize the domain gap.

Main Results:

• The proposed source-free UDA method achieves performance comparable or superior to traditional UDA methods.
• Adapted models demonstrate significantly improved robustness when subjected to input image perturbations.
• Effectively reduces the domain gap between source and target data distributions.

Conclusions:

• The developed source-free UDA approach offers a viable and effective solution for domain adaptation without source data.
• The method enhances model adaptability and robustness, particularly in challenging scenarios with image variations.