TsCANet: Three-stream contrastive adaptive network for cross-domain few-shot learning

Cross-domain few-shot learning, which aims to solve the problem of domain gap in few-shot learning, has recently received more and more attention. Specifically, when there are great differences between the source domain and the target domain involved in few-shot learning, the performance will fall off a cliff and it is even difficult to train. Therefore, this paper explores a simple, effective and novel method to deal with domain gaps. Firstly, the pre-trained model is obtained by using the labeled data in the source domain. Next, the two-stage adaptive training mainly consists of unlabeled data in the target domain, pseudo-unlabeled data and labeled data in the source domain as the third-stream input, so that the network can gradually adapt to the data in target domain and mitigate the adverse effects caused by the domain gap. Finally, the proposed network can be quickly applied to the tasks to be solved. Through the observation of experimental results, the designed approach can achieve better performance than the existing comparison methods on the standard benchmark of cross-domain few-shot learning. Further analysis reveals the tradeoff between using data in source domain and target domain for cross-domain few-shot learning.