Few-shot classification with task-adaptive semantic feature learning

Meihong Pan, Hongyi Xin, Chun-Qiu Xia, Hong-Bin Shen

Abstract: Highlights•A task-adaptive semantic feature learner is proposed to predict semantic features from images for the query samples, which alleviates the information imbalance between the support and query sets and improves the discrimination of the query features.•The visual and semantic features are learned separately without collapse into a common feature space, which maintains specific information of different modalities.•Some hard tasks with samples having similar background and interference are constructed to prove the efficacy of TasNet in generating discriminative features from hard samples.

External IDs:dblp:journals/pr/PanXXS23