Pretext Tasks in Test Time Adaptation Under Distribution Shifts-A Survey and Future Directions

The performance of well-trained models deteriorates significantly under distribution shifts between training and test datasets. In response, various test time adaptive methods have been proposed to narrow domain gaps by capturing distribution cues from test samples. Notably, pretext task-based test time adaptive models exhibit promising performance, as they do not require target test annotations and separate training and testing stages, by leveraging well-designed pretext tasks, enabling effective adaptation at test time. Moreover, to accommodate diverse scenarios, task-specific pretext tasks are proposed to improve adaptive performance. Currently, a review has provided a comprehensive overview of test time adaptive methods. Nevertheless, there remains a notable gap in detailed surveys of pretext tasks employed in test time adaptation. To narrow this gap, this paper presents a survey of pretext tasks employed in test time adaptive models. We begin by providing an overview of test time adaptive methods, followed by giving a concise review of pretext tasks used in common scenarios, with a comparison to those used in test time adaptation scenarios. Subsequently, we delve into pretext tasks employed in various test time adaptation scenarios, exploring their characteristics, strengths, and limitations. Lastly, we conduct an empirical analysis with various pretext tasks in a digit prediction task, and subsequently conclude with a discussion of potential directions for future research.