Unsupervised domain adaptation (UDA) aims to learn a prediction model for the target domain given labeled source data and unlabeled target data. Impressive progress has been made by adversarial learning-based methods that align distributions across domains through deceiving a domain discriminator network. However, these methods only try to align two domains and neglect the boundaries between classes, which may lead to false alignment and poor generalization performance. In contrast, consistencyenforcing methods exploit the target data posterior distribution to make the target features far away from decision boundaries. Despite their efficacy, these approaches require additional intensity augmentation to align distributions when encountering datasets with large domain discrepancy. To solve the above problems, we propose a novel UDA method that unifies the adversarial learning-based method and consistencyenforcing method together to take both domain alignment and boundaries between classes into consideration. In addition to the supervised classification on the source domain and the adversarial domain adaptation, we introduce interpolation consistency into the UDA task. To be specific, we first construct robust and informative pseudo labels for target samples, and then we encourage the prediction at an interpolation of unlabeled target samples to be consistent with the interpolation of the pseudo labels of these samples. The extensive empirical results demonstrate that our method achieves state-of-the-art results on both digit classification and object recognition tasks.

Introduction

Deep learning approaches have achieved remarkable success in various computer vision tasks and applications. However, these achievements often rely on large-scale labeled datasets. In many cases, the collection and annotation of training data on novel domains are extremely expensive or sometimes impossible. Hence, there is a strong motivation to train a good classification model for a target domain by using readilyavailable annotated data from a source domain with a different distribution. However, this attractive transfer learning paradigm suffers from the data shift problem [1], which is a huge challenge for adapting classification models to the target domain. Learning a classifier under data shift between the labeled source domain and unlabeled target domain is known as unsupervised domain adaptation (UDA) [2]. Many UDA approaches directly align the marginal distributions across domains to bridge the domain gap [3]–[۹]. Notably, approaches based on adversarial learning [3], [4] divide the base model into a feature extractor G and a task-specific classifier C, and add a domain discriminator D. The domain discriminator D takes the features extracted by G and predicts which domains the features come from. The feature extractor G is learned to extract domain-invariant feature representations by deceiving the domain discriminator. Domain alignment is expected when the adversarial training reaches an equilibrium. However, these approaches may fail to create discriminative features because they do not consider the decision boundary.