Towards Corruption-Agnostic Robust Domain Adaptation

Great progress has been achieved in domain adaptation decades. Existing works are always based on an ideal assumption that testing target domains independent and identically distributed with training domains. However, due to unpredictable corruptions (e.g., noise blur) real data, such as web images real-world object detection, methods increasingly required be corruption robust We investigate a new task, corruption-agnostic (CRDA), accurate original data against unavailable-for-training This task is non-trivial the large discrepancy unsupervised observe simple combinations of popular robustness have suboptimal CRDA results. propose approach two technical insights into CRDA, follows: (1) easy-to-plug module called generator (DDG) generates samples enlarge mimic corruptions; (2) but effective teacher-student scheme contrastive loss enhance constraints Experiments verify DDG maintains or even improves its performance achieves better than baselines. Our code available at: https://github.com/YifanXu74/CRDA .