Cross-domain self-supervised few-shot learning via multiple crops with teacher-student network

Most few-shot learning(FSL) methods rely on a pre-trained network on a large annotated base dataset with a feature distribution similar to that of the target domain. Conventional transfer learning and traditional few-shot learning methods are ineffective when there is a large gap between the source and target domain. We propose a simple teacher-student network solution to facilitate unlabeled images from the target domain to alleviate domain gap. We impose a self-supervised loss by calculating predictions from large crops of the unannotated samples of target domain using a teacher network and matching them with small crops of the same images from a student network. Furthermore, we design a novel contrastive loss for large crops to sufficiently utilize the self-supervised information of unlabeled images on target domain for the model training. The feature representation can be easily generalized to the target domain without the pretraining phase on target-specific classes. The accuracies of our model are 23.61 +/- 0.42, 33.87 +/- 0.59, 63.21 +/- 0.88, 74.36 +/- 0.88 on ChestX, ISIC, EuroSAT, and CropDisease datasets for the 1-shot scenario respectively. Extensive experiments show that the proposed method achieves competitive performance on the challenging cross-domain FSL image classification.