Random Erasing Data Augmentation

In this paper, we introduce Random Erasing, a simple yet effective data augmentation techniques for training the convolutional neural network (CNN). In training phase, Random Erasing randomly selects a rectangle region in an image, and erases its pixels with random values. In this process, training images with various levels of occlusion are generated, which reduce the risk of network overfitting and make the model robust to occlusion. Random Erasing is parameter learning free, easy to implement, and can be integrated into most of the CNN-based recognition models. Albeit simple, Random Erasing yields consistent improvement in image classification, object detection and person re-identification (re-ID). For image classification, our method improves WRN-28-10: top-1 error rate from 3.72% to 3.08% on CIFAR10, and from 18.68% to 17.65% on CIFAR100. For object detection on PASCAL VOC 2007, Random Erasing improves Fast-RCNN from 74.8% to 76.2% in mAP. For person re-ID, when using Random Erasing in recent deep models, we achieve the state-of-the-art accuracy: the rank-1 accuracy is 89.13% for Market-1501, 84.02% for DukeMTMC-reID, and 63.93% for CUHK03 under the new evaluation protocol.