Addressing distribution shift in computer vision

Abstract

however, they often suffer from notable performance degradation when confronted with distribution shifts between the train set and the test set. Addressing this challenge without the need for additional data collection has become a recent research focus. A primary hurdle in handling such shifts is the dataset bias, where models overly rely on the unwanted correlation between peripheral attributes and labels. This bias can lead models to learn irrelevant features, hindering their ability to generalize to various data distributions. Another challenge is the domain shift, which encompasses differences in style, object sizes, or sources of datasets. Among various techniques for addressing the domain shift, test-time adaptation (TTA) has gained traction for its practicality in mitigating these shifts. This thesis mainly tackles such two challenges in computer vision and further suggests the future work direction of addressing distribution shifts.

In computer vision, deep neural networks have made significant progresses