Inferring Latent Domains for Unsupervised Deep Domain Adaptation

Unsupervised Domain Adaptation (UDA) refers to the problem of learning a model in target domain where labeled data are not available by leveraging information from annotated source domain. Most deep UDA approaches operate single-source, single-target scenario, i.e., they assume that and samples arise single distribution. However, practice most datasets can be regarded as mixtures multiple domains. In these cases, exploiting traditional methods for classification models may lead poor results. Furthermore, it is often difficult provide labels all points, i.e. latent domains should automatically discovered. This paper introduces novel architecture which addresses discovering visual this learn robust classifiers. Specifically, our based on two main components, side branch computes assignment each sample its layers exploit membership appropriately align distribution CNN internal feature representations reference We evaluate approach publicly benchmarks, showing outperforms state-of-the-art adaptation methods.