Learn to aggregate global and local representations for few-shot learning

Few-shot learning aims to train recognition models to learn new object categories from limited training examples. Recent metric-learning based methods have made significant progress. Most of these methods rely on a single similarity metric at a global or local level. However, classifying samples using multiple similarity metrics at different levels simultaneously can produce a better similarity measure and more discriminative feature maps. Therefore, in this paper, a novel method called Learn to Aggregate Global and Local Representations for Few-shot Learning is introduced. Our proposed method embeds the support images and the query images. Then, it calculates four distinct similarity metrics between representations at global and local levels. Finally, the calculated similarities are combined and fed to a fusion module to obtain a final similarity score. Extensive experiments demonstrate that our method achieves state-of-the-art results on popular benchmarks. Particularly, AGLRs outperforms DN4 with a margin of ≈ 3 − 4% on the miniImageNet dataset.