Cross-Domain Few-Shot Classification via Learned Feature-Wise Transformation

Few-shot classification aims to recognize novel categories with only few\nlabeled images in each class. Existing metric-based few-shot classification\nalgorithms predict categories by comparing the feature embeddings of query\nimages with those from a few labeled images (support examples) using a learned\nmetric function. While promising performance has been demonstrated, these\nmethods often fail to generalize to unseen domains due to large discrepancy of\nthe feature distribution across domains. In this work, we address the problem\nof few-shot classification under domain shifts for metric-based methods. Our\ncore idea is to use feature-wise transformation layers for augmenting the image\nfeatures using affine transforms to simulate various feature distributions\nunder different domains in the training stage. To capture variations of the\nfeature distributions under different domains, we further apply a\nlearning-to-learn approach to search for the hyper-parameters of the\nfeature-wise transformation layers. We conduct extensive experiments and\nablation studies under the domain generalization setting using five few-shot\nclassification datasets: mini-ImageNet, CUB, Cars, Places, and Plantae.\nExperimental results demonstrate that the proposed feature-wise transformation\nlayer is applicable to various metric-based models, and provides consistent\nimprovements on the few-shot classification performance under domain shift.\n