Source-Free Unsupervised Domain Adaptation through Trust-Guided Partitioning and Worst-Case Aligning

Abstract

In source-free unsupervised domain adaptation (SFUDA) tasks, adapting to the target domain without directly accessing the source domain data and relying solely on a pre-trained source domain model and the target domain data is a common challenge. Existing approaches often rely on pseudo-labeling techniques for intra-class clustering to achieve global alignment of classes. However, the presence of noise can lead to incorrect clustering results. In this paper, we introduce a novel approach referred to as Trust-guided Partitioning and Worst-case Aligning (TPWA). We assess the reliability of pseudo-labels by computing the similarity difference between the class centers corresponding to the pseudo-labels and the centers of the most similar classes. Based on this, we perform partitioning and then conduct intra-class clustering only on high-trustworthy samples. We also train a worst-case classifier to predict correctly on high-trustworthy samples and make as many mistakes as possible on low-trustworthy samples, and then adversarially trains feature extractors to align low-trustworthy samples to high-trustworthy samples. This approach leverages the structural information present in the high-trustworthy sample set, improving the robustness of the adaptation. Additionally, we also consider enforcing prediction consistency among neighboring samples to further constrain the pseudo-labels. Extensive experiments demonstrate the superiority of our method in SFUDA tasks.