Random Erasing Data Augmentation

The paper introduces Random Erasing, a new data augmentation method for training convolutional neural networks. During training, Random Erasing randomly selects a rectangle region within an image and erases its pixels by replacing them with random values, thereby generating training images that exhibit various levels of occlusion. This process is designed to reduce the risk of overfitting and to make the trained model more robust to occlusion, and the technique is parameter learning free, easy to implement, and can be integrated with most CNN-based recognition models.

Although simple, Random Erasing is complementary to commonly used data augmentation techniques such as random cropping and flipping, and it yields consistent improvement over strong baselines across multiple tasks, including image classification, object detection, and person re-identification. Because it introduces no learnable parameters and is trivial to plug into existing training pipelines, it offers an easy and general way to improve robustness and generalization, and the authors make their code publicly available.