Distribution Estimation Based Pseudo-Feature Library Generation For Few-Shot Image Classification

Jiacheng Zhang, Weishuo Liu, Zhicheng Zhao, Fei Su

Abstract: Due to the high cost of labeled data acquisition, few-shot learning has attracted great attention in recent years. The biased estimation of class distribution from a few labeled samples hinders the model’s performance. Some existing methods generate samples or features by a learning module or network. In this paper, a distribution-based pseudo-feature library generation method is proposed, and it follows a simple distribution modeling hypothesis. The base class features is introduced to better estimate the novel class distribution. Furthermore, a patch-level pseudo-feature library is adversarially generated to reinforce the training of the classifier. The proposed method significantly improves the model performance for the few-shot image classification task without introducing additional training parameters. Our method ranks first in the ICME 2021 Few-Shot Learning for Vehicle Footprint Recognition Challenge, demonstrating its effectiveness.