A Coarse-to-Fine Auto-Sampler For Long-tailed Image Recognition

Tong Wu, Hao Li

Keywords: Reinforcement Learning, AutoML

Abstract: The long-tail distributed data in the real world has always been a great challenge for deep learning. Current approaches to mitigate the long-tail issue include re-sampling and data augmentation. However, the hand-crafted re-sampling and augmentation strategies are sub-optimal. In this paper, we propose a coarse-to-fine auto-sampler that re-samples and augments imbalanced data automatically using reinforcement learning. The sampler consists of two parts: a patch sampler for augmenting data during representation learning, and a class-wise weighted sampler for classifier learning. Experiments on standard long-tailed datasets including CIFAR-10/100-LT and ImageNet-LTT shows the effectiveness of our method. We also conduct visualization to better demonstrate the policy of our agents. Our code will be made public before the final revision. Our video is available at https://drive.google.com/file/d/ 1oamWNap3rC3Ulyd5jSs1ZrZIQiBB-iR1/view?usp=sharing.