Where to Cut and Paste: Data Regularization with Selective Features

Abstract

Deep convolutional neural networks are continually evolving through various effective training methods such as data augmentation. Among data augmentation methods, regional dropout or replacement strategies such as [3], [4], [5] have been proved effective in recognition and localization performance. However, such methods suffer from unintended content corruption like informative pixel loss. For example, cutting and pasting a random patch may consist of areas that are not important and even if a new cutout patch consists of informative pixels, it could be pasted at useful locations of input covering the interest of the object. Therefore, this operation can cause too much or meaningless regularization. Motivated by this, we propose a new data augmentation method strategy, called FocusMix, which exploits informative pixels based on proper sampling techniques. Through experiments, we analyzed and compared various data augmentation methods to provide improvements and effectiveness of FocusMix. Finally, we have shown that FocusMix results in improvements in performance compared to other data augmentation methods.