Pairwise Similarity for Domain Adaptation

Abstract

The application of the domain adaptation technique enables the resolution of classification challenges in an unlabeled target domain by leveraging the labeled information from source domains. Nevertheless, prior approaches to domain adaptation have primarily concentrated on global domain adaptation, neglecting the inclusion of class-specific data, thereby resulting in substandard performance when transferring knowledge. Recently, there has been an increasing interest in employing similarity measures for inter-domain and intra-domain alignment, focusing more meticulously on the fine-grained information of samples. However, due to the inherent complexity of multi-class image data, it has been challenging for existing studies to address inter-domain and intra-domain differences by assigning weights through a single similarity measure, thus failing to effectively mitigate the classification issues of dissimilar samples. To address this challenge, we have introduced a new method, named Pairwise Similarity for Domain Adaptation (PSDA), as a solution to this problem. Specifically, we allocate different similarity weights under various similarity measures to enhance both inter-domain and intra-domain alignment. By integrating the characteristics of different similarity weights, the network model is able to complementarily learn feature information under both modes, effectively improving the model’s adaptability. Comprehensive experiments conducted on four widely used public datasets have demonstrated that our method surpasses many state-of-the-art domain adaptation approaches, showing significant improvements over many advanced methods.

Keywords

• Domain adaptation
• inter-domain and intra-domain alignment
• similarity measure