FDDS: Feature Disentangling and Domain Shifting for Domain Adaptation

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