Deep adaptation relation networks for across domain classification

Deep learning has achieved great success in visual recognition tasks which requires a great of annotated samples. However, annotated samples are not available in many situations. Domain adaptation improves the performance on an unlabeled target domain by utilizing the knowledge learned from a related source domain. How to match the domain distributions and transfer the source model to target ones when the labeled target samples cannot reflect the whole target domain distribution will be challenging tasks. In this paper, we provide a relation metric strategy which can learn a deep distance metric to compare a small set within episodes, so we can exploit the unlabeled data from target domain with a few-shot set. Furthermore, we use conditional distributions adaptation to close the difference between two domains by build pseudo labels. The comparative experiments demonstrate that the proposed network works better than previous methods on the standard domain adaptation benchmarks.