tSF: Transformer-based Semantic Filter for Few-Shot Learning

• DOI: 10.48550/arxiv.2211.00868
• Publication Date: 2022-01-01
• Authors: Jinxiang Lai; Siqian Yang; Wenlong Liu; Yi Zeng; Zhongyi Huang; Wenlong Wu; Jun Liu; Bin-Bin Gao; Chengjie Wang
• Keywords: Discriminative model; Feature Learning

Few-Shot Learning (FSL) alleviates the data shortage challenge via embedding discriminative target-aware features among plenty seen (base) and few unseen (novel) labeled samples. Most feature modules in recent FSL methods are specially designed for corresponding learning tasks (e.g., classification, segmentation, object detection), which limits utility of features. To this end, we propose a light universal module named transformer-based Semantic Filter (tSF), can be applied different tasks. The proposed tSF redesigns inputs structure by semantic filter, not only embeds knowledge from whole base set to novel but also filters target category. Furthermore, parameters is equal half standard transformer block (less than 1M). In experiments, our able boost performances classic few-shot (about 2% improvement), especially outperforms state-of-the-arts on multiple benchmark datasets classification task.