FDDS: Feature Disentangling and Domain Shifting for Domain Adaptation

Abstract

Domain adaptation is a learning strategy that aims to improve the performance of models in the current field by leveraging similar domain information. In order to analyze the effects of feature disentangling on domain adaptation and evaluate a model’s suitability in the original scene, we present a method called feature disentangling and domain shifting (FDDS) for domain adaptation. FDDS utilizes sample information from both the source and target domains, employing a non-linear disentangling approach and incorporating learnable weights to dynamically separate content and style features. Additionally, we introduce a lightweight component known as the domain shifter into the network architecture. This component allows for classification performance to be maintained in both the source and target domains while consuming moderate overhead. The domain shifter uses the attention mechanism to enhance the ability to extract network features. Extensive experiments demonstrated that FDDS can effectively disentangle features with clear feature separation boundaries while maintaining the classification ability of the model in the source domain. Under the same conditions, we evaluated FDDS and advanced algorithms on digital and road scene datasets. In the 19 classification tasks for road scenes, FDDS outperformed the competition in 11 categories, particularly showcasing a remarkable 2.7% enhancement in the accuracy of the bicycle label. These comparative results highlight the advantages of FDDS in achieving high accuracy in the target domain.