UC San Diego

Recent advancements in visual categorization have led to significant improvements across various applications, but these models still struggle to generalize effectively to under-represented regions and demographics, directly impacting the fairness and inclusivity of computer vision systems. While domain adaptation has been proposed as a solution to bridge domain gaps using unlabeled data, its effectiveness in handling complex distribution shifts remains insufficiently explored. This dissertation explores efforts to enhance robustness and transferability in computer vision through domain adaptation. First, we introduce an efficient mechanism to scale domain adaptation to categorization tasks with hundreds of classes. Next, we introduce GeoNet, a dataset designed to benchmark and analyze geographical disparities in visual categorization tasks. Later, we present our work on using language as a powerful tool to guide the learning of transferable representations across different domains in images and videos, followed by UDABench, a new unified framework aimed at standardizing the training and evaluation of domain adaptation algorithms along with share key insights from this framework. Lastly, we identify significant open research questions that could further advance the concepts discussed in this dissertation.