Adversarially Trained Object Detector for Unsupervised Domain Adaptation

Unsupervised domain adaptation, which involves transferring knowledge from a label-rich source to an unlabeled target domain, can be used substantially reduce annotation costs in the field of object detection. In this study, we demonstrate that adversarial training employed as new approach for unsupervised adaptation. Specifically, establish adversarially trained detectors achieve improved detection performance domains are significantly shifted domains. This phenomenon is attributed fact extract robust features alignment with human perception and worth across while discarding domain-specific non-robust features. addition, propose method combines feature ensure domain. We conduct experiments on four benchmark datasets confirm effectiveness our proposed large shifts real artistic images. Compared baseline models, improve mean average precision by up 7.7%, further 11.8% when alignments incorporated. Although degrades small shifts, quantification shift based Frechet distance allows us determine whether should conducted.