Class-Level Adaptation Network with Self Training for Unsupervised Domain Adaptation.

Deep learning has been widely used in various tasks. However, in real-world scenarios, obtaining various dataset with labels is time-consuming and difficult. Most models are trained in simulated scenarios, and such models degrade in real-world scenarios. Unsupervised domain adaptation is a branch of transfer learning, which utilizes a large number of labeled source domain data to improve the performance of the model in the target domain with limited or missing labels through knowledge transfer. However, most of the previous work neglected the category information when aligning the distribution between source and target domains, which led to the emergence of negative transfer. To address this problem, we propose the class-level adaptation network (CLAN) optimizing a novel metric which makes the centers of each class in source and target domains are close. Specifically, the class center of the source domain is generated by the labels of the source samples, while the class center of the target domain without any label is generated by the high confidence pseudo labels of the target samples. Technically, CLAN matches each target sample to the nearest center in the source domain and assigns an example a high confidence pseudo label by considering a threshold. Extensive experiments indicate that the combination of the aforementioned two models can achieve state-of-the-art performance on the Office-31 and digital domain adaptation benchmarks.