Data Augmentation Using Mixup and Random Erasing

Abstract

Deep convolutional neural networks show excellent performance on computer vision tasks. However, these networks rely heavily on large-scale datasets to avoid overfitting. Un-fortunately, except for some datasets used for classical tasks, only small-scale datasets can be acquired in many applications. Data augmentation is a commonly used approach to extend the dataset scale and take advantage of the capabilities of large-scale datasets. Based on Mixup and random erasing, this paper proposes two different combinations of these two methods, namely RSM and RDM, to compensate their respective shortcomings. The RSM method mix up two original images before erasing randomly selected region, while the RDM method performs in opposite order. The two proposed methods are evaluated extensively for object detection and image classification on various datasets. The experimental results show that RSM and RDM achieve over 1% and 1.5% improvements for the detection of small-scale objects and the image classification, respectively, compared to Mixup and random erasing.