English

Prototype Completion with Primitive Knowledge for Few-Shot Learning

Computer Vision and Pattern Recognition 2021-06-25 v6

Abstract

Few-shot learning is a challenging task, which aims to learn a classifier for novel classes with few examples. Pre-training based meta-learning methods effectively tackle the problem by pre-training a feature extractor and then fine-tuning it through the nearest centroid based meta-learning. However, results show that the fine-tuning step makes very marginal improvements. In this paper, 1) we figure out the key reason, i.e., in the pre-trained feature space, the base classes already form compact clusters while novel classes spread as groups with large variances, which implies that fine-tuning the feature extractor is less meaningful; 2) instead of fine-tuning the feature extractor, we focus on estimating more representative prototypes during meta-learning. Consequently, we propose a novel prototype completion based meta-learning framework. This framework first introduces primitive knowledge (i.e., class-level part or attribute annotations) and extracts representative attribute features as priors. Then, we design a prototype completion network to learn to complete prototypes with these priors. To avoid the prototype completion error caused by primitive knowledge noises or class differences, we further develop a Gaussian based prototype fusion strategy that combines the mean-based and completed prototypes by exploiting the unlabeled samples. Extensive experiments show that our method: (i) can obtain more accurate prototypes; (ii) outperforms state-of-the-art techniques by 2% - 9% in terms of classification accuracy. Our code is available online.

Keywords

Cite

@article{arxiv.2009.04960,
  title  = {Prototype Completion with Primitive Knowledge for Few-Shot Learning},
  author = {Baoquan Zhang and Xutao Li and Yunming Ye and Zhichao Huang and Lisai Zhang},
  journal= {arXiv preprint arXiv:2009.04960},
  year   = {2021}
}

Comments

Accepted by CVPR2021

R2 v1 2026-06-23T18:27:02.057Z