Few-shot learning with long-tailed labels

Hongliang Zhang, Shuo Chen, Lei Luo, Jiang Yang

Abstract: Highlights•We propose a new problem setting termed FSL-LTL to consider a frequently occurring practical issue in which the class labels are long-tailed.•We build a novel two-stage training framework called RCE to solve this new problem. It consists of a pre-training phase based on contrastive learning and a semi-supervised learning phase based on weighted sampling.•Experimental results demonstrate the superiority of our proposed method over mainstream FSL methods and semi-supervised learning methods on FSL-LTL and the results suggest that our RCE is a promising solution for this new problem setting.