DEEP VISUAL DOMAIN ADAPTATION IN THE WILD

Abstract

This thesis presents a systematic study on deep visual domain adaptation, with an emphasis on real-world applications. The research explores three main aspects: developing efficient unsupervised domain adaptation (UDA) methods, exploring practical UDA configurations, and investigating effective unsupervised validation methods.

The first work proposes the NOUN and PRONOUN methods, which are simple yet effective conditional domain adversarial training approaches for UDA. These methods outperform previous domain adversarial training methods in both classification and segmentation tasks.

The second work introduces the UMAD framework, which handles general open-set UDA scenarios without requiring access to source data or prior knowledge about category overlap. UMAD achieves comparable performance to UDA methods that rely on source data.

Finally, a novel and simple validation method MixVal is proposed. MixVal is a generic validation method that utilizes unlabeled target data and synthesizes new target samples using mixup. It outperforms other methods across different validation settings.