Abstract: Generalization under distribution shifts has been a great challenge in computer vision. The prevailing practice of directly employing the one-hot labels as the training targets in domain ...