Wave-SAN: Wavelet based Style Augmentation Network for Cross-Domain Few-Shot Learning

Previous few-shot learning (FSL) works mostly are limited to natural images\nof general concepts and categories. These works assume very high visual\nsimilarity between the source and target classes. In contrast, the recently\nproposed cross-domain few-shot learning (CD-FSL) aims at transferring knowledge\nfrom general nature images of many labeled examples to novel domain-specific\ntarget categories of only a few labeled examples. The key challenge of CD-FSL\nlies in the huge data shift between source and target domains, which is\ntypically in the form of totally different visual styles. This makes it very\nnontrivial to directly extend the classical FSL methods to address the CD-FSL\ntask. To this end, this paper studies the problem of CD-FSL by spanning the\nstyle distributions of the source dataset. Particularly, wavelet transform is\nintroduced to enable the decomposition of visual representations into\nlow-frequency components such as shape and style and high-frequency components\ne.g., texture. To make our model robust to visual styles, the source images are\naugmented by swapping the styles of their low-frequency components with each\nother. We propose a novel Style Augmentation (StyleAug) module to implement\nthis idea. Furthermore, we present a Self-Supervised Learning (SSL) module to\nensure the predictions of style-augmented images are semantically similar to\nthe unchanged ones. This avoids the potential semantic drift problem in\nexchanging the styles. Extensive experiments on two CD-FSL benchmarks show the\neffectiveness of our method. Our codes and models will be released.\n