DAC: Disentanglement-and-Calibration Module for Cross-Domain Few-Shot Classification

Abstract

Cross-domain few-shot classification (CD-FSC) aims to develop few-shot classification models trained on seen domains but tested on unseen domains. However, the cross-domain setup poses a challenge in the form of domain shift between the training and testing domains. Previous research has demonstrated that the encoder can disentangle features into domain-shared and domain-specific features. However, poorly estimated domain-specific features can lead to inadequate generalization on the unseen domain. This paper proposes a disentanglement-and-calibration module (DAC) to address this issue. The disentanglement component separates the features into domain-shared and domain-specific features, while the calibration component corrects the domain-specific features. We demonstrate that the DAC module can significantly enhance the generalization capability of several baseline methods. Furthermore, we show that MatchingNet with the DAC module outperforms existing state-of-the-art methods by 10%-11% when trained on mini-ImageNet, CUB-200, Cars196, Places365 and tested on Plantae dataset.

Keywords

• Cross-domain few-shot classification
• disentanglement
• domain shift
• representation learning