ESPT: A Self-Supervised Episodic Spatial Pretext Task for Improving Few-Shot Learning

Self-supervised learning (SSL) techniques have recently been integrated into the few-shot (FSL) framework and shown promising results in improving image classification performance. However, existing SSL approaches used FSL typically seek supervision signals from global embedding of every single image. Therefore, during episodic training FSL, these methods cannot capture fully utilize local visual information samples data structure whole episode, which are beneficial to FSL. To this end, we propose augment objective with a novel self-supervised Episodic Spatial Pretext Task (ESPT). Specifically, for each generate its corresponding transformed episode by applying random geometric transformation all images it. Based on these, our ESPT is defined as maximizing spatial relationship consistency between original one. With definition, ESPT-augmented promotes more transferable feature representations that features different their inter-relational structural input thus enabling model generalize better new categories only few samples. Extensive experiments indicate method achieves state-of-the-art performance three mainstay benchmark datasets. The source code will be available at: https://github.com/Whut-YiRong/ESPT.