Class Activation Map based Random Erasing for Data Augmentation

Random erasing offers various levels of occlusion for data augmentation. However, due to its uniform distribution of random selection, it sometimes occludes regions that are unrelated to the object of interest. In this paper, we propose a novel method that utilizes Gradient Weighted Class Activation Mapping (Grad-CAM) for estimating the location of the object of interest and selectively erasing the surrounding areas. By utilizing Grad-CAM, we improve random erasing for CNN models without requiring additional modules or architectural changes. We generate Grad-CAM after the intermediate epochs where CNN models have sufficient representational power for the training data. The hyperparameter that restrict the erasing to the vicinity of the object is set based on Grad-CAM, and experiments were conducted accordingly. As a result of our experiments, we observed a 0.33% decrease in error-rate for image classification tasks using ResNet-20 on the CIFAR-10 dataset.