Class Consistency Driven Unsupervised Deep Adversarial Domain Adaptation

In unsupervised deep domain adaptation (DA), the use of adversarial domain classifiers is popular in learning a shared feature space which reduces the distributions gap for a pair of source (with training data) and target (with only test data) domains. In the new space, a classifier trained on source training data is expected to generalize well for the target domain samples. We hypothesize that such a feature space obtained by aligning the domains globally ignores the category level feature distributions. This, in turn, leads to erroneous mapping for fine-grained classes. Besides, the discriminativeness of the shared space is not explicitly addressed. In order to resolve both the issues, we propose a novel adversarial approach which judiciously refines the space learned by the domain classifier by incorporating class level information. We follow an ensemble classifiers based approach to model the source domain and introduce a novel consistency constrain on the classifier's outcomes when evaluated on a held-out set of target domain samples. We further leverage the ensemble learning strategy during the inference, as opposed to the existing single classifier based methods. We find that our deep DA model is capable of producing a compact and better domain aligned feature space. Experimental results obtained on the Office-Home, Office-CalTech, MNIST-ASPS, and a remote sensing dataset confirm the superiority of the proposed approach.