Self-randomized focuses effectively boost metric-based few-shot classifiers

Zhen Li, Zhongyuan Liu, Dongliang Chang, Aneeshan Sain, Xiaoxu Li, Zhanyu Ma, Jing-Hao Xue, Yi-Zhe Song

Abstract: Highlights•Data augmentation can motivate the model to focus on more discriminative regions and alleviate overfitting.•Data augmentation methods suffer from the problem of removing or blurring the discriminative parts.•Two simple and effective data augmentation methods are proposed.•The proposed methods significantly improves the performance of the baseline methods.