Adversarially Trained Object Detector for Unsupervised Domain Adaptation

Abstract

Unsupervised domain adaptation, which involves transferring knowledge from a\nlabel-rich source domain to an unlabeled target domain, can be used to\nsubstantially reduce annotation costs in the field of object detection. In this\nstudy, we demonstrate that adversarial training in the source domain can be\nemployed as a new approach for unsupervised domain adaptation. Specifically, we\nestablish that adversarially trained detectors achieve improved detection\nperformance in target domains that are significantly shifted from source\ndomains. This phenomenon is attributed to the fact that adversarially trained\ndetectors can be used to extract robust features that are in alignment with\nhuman perception and worth transferring across domains while discarding\ndomain-specific non-robust features. In addition, we propose a method that\ncombines adversarial training and feature alignment to ensure the improved\nalignment of robust features with the target domain. We conduct experiments on\nfour benchmark datasets and confirm the effectiveness of our proposed approach\non large domain shifts from real to artistic images. Compared to the baseline\nmodels, the adversarially trained detectors improve the mean average precision\nby up to 7.7%, and further by up to 11.8% when feature alignments are\nincorporated. Although our method degrades performance for small domain shifts,\nquantification of the domain shift based on the Frechet distance allows us to\ndetermine whether adversarial training should be conducted.\n