Abstract: In the real world, there often exist distribution shifts in the data. Domain Adaptation overcomes this issue with an aim to transfer a source domain model to a related but different target domain. Domain adaptation is a critical task in many visual applications such as object recognition, semantic segmentation, stylized image generation, etc.

Despite its popularity, domain adaptation is still a challenging task, especially when the domain gap is large. In this thesis proposal, we develop several efficient domain adaptation approaches to improve model performance on the target domain. First, inspired by the large-scale pretraining of visual Transformers, we explore Transformer-based domain adaptation for stronger feature representation and design a safe training mechanism to avoid model collapse. Then, we focus on the penultimate activations of target domain data, and propose an adversarial training strategy to enhance model prediction confidences. Next, we propose to rectify model predictions using prior knowledge of target domain label distribution. A novel Knowledge-guided Unsupervised Domain adaptation paradigm is proposed. Finally, we step into the task of Active Domain Adaptation. We propose a novel active selection criterion based on the local context, and devise a progressive augmentation module to better utilize queried target data.

To conclude this thesis, we plan to investigate the challenging task of zero-shot and open-class adaptation. Large-scale visual-language models like CLIP make it possible to transfer models to unseen classes using language descriptions. We will briefly talk about existing works and future research plans.