Adaptive set-level metric for few-Shot image classification.

Abstract

Few-shot image classification aims to learn a classifier from limited labeled data. Though the existing methods have achieved significant improvement, they are still challenging to accurately differentiate image categories between the confused support and query samples. We observed that the objects belong to same category may exhibit significant image-level appearance difference between the support and query, while the objects belong to different categories may have the similar appearance on them. To this end, we propose to represent each image using a set of feature embeddings instead of only one image-level feature, by which allowing the network to aggregate more rich and useful features from different views of the image. Furthermore, we propose a set-based metric approach with dynamic self-adapting weights mechanism to measure the similarity between the two sets of image embeddings (the query set and the support set). Meanwhile, to improve the accuracy of the dynamic self-adapting weights, we introduced primitive knowledge (i.e., class-level part or attribute annotations) as a priori knowledge to adjust these weights. The experimental results demonstrate the effectiveness of the proposed method, achieving state-of-the-art performance in miniImageNet, tieredImageNet, and CUB dataset (over 0.62 %, 1.69 %, 1.09 % respectively improvement compared to the best competing method).