TGDM: Target Guided Dynamic Mixup for Cross-Domain Few-Shot Learning

Abstract

Given sufficient training data on the source domain, cross-domain few-shot\nlearning (CD-FSL) aims at recognizing new classes with a small number of\nlabeled examples on the target domain. The key to addressing CD-FSL is to\nnarrow the domain gap and transferring knowledge of a network trained on the\nsource domain to the target domain. To help knowledge transfer, this paper\nintroduces an intermediate domain generated by mixing images in the source and\nthe target domain. Specifically, to generate the optimal intermediate domain\nfor different target data, we propose a novel target guided dynamic mixup\n(TGDM) framework that leverages the target data to guide the generation of\nmixed images via dynamic mixup. The proposed TGDM framework contains a Mixup-3T\nnetwork for learning classifiers and a dynamic ratio generation network (DRGN)\nfor learning the optimal mix ratio. To better transfer the knowledge, the\nproposed Mixup-3T network contains three branches with shared parameters for\nclassifying classes in the source domain, target domain, and intermediate\ndomain. To generate the optimal intermediate domain, the DRGN learns to\ngenerate an optimal mix ratio according to the performance on auxiliary target\ndata. Then, the whole TGDM framework is trained via bi-level meta-learning so\nthat TGDM can rectify itself to achieve optimal performance on target data.\nExtensive experimental results on several benchmark datasets verify the\neffectiveness of our method.\n