DropMask: A data augmentation method for convolutional networks

Abstract

Convolutional neural networks can learn a powerful feature space and play a great role on the promotion of autonomous driving. On limited datasets, deep neural networks often overfit. Properly adding noise and regularization during training can alleviate this. DropBlock randomly removes blocks on the feature map, however, such randomness may lead to complete removal of objects due to excessive removal of one or a few blocks, as well as contextual information. To solve this problem, we propose a regularization method (DropMask), which strikes a reasonable balance between deletion and retention in the block to be deleted on the feature map. That will avoid excessive removal of contiguous blocks so that improve the accuracy and robustness of model. After extensive experiments, it has been shown that the DropMask proposed in the paper outperforms DropBlock on neural networks. On CIFAR-I0 classification, ResNet-18 architecture with DropMask achieves 95.34% on accuracy, 1.88% improvement over the baseline. On KITTI 2D detection task, Yolov5s with DropMask improves mAP from 77.6% to 79.2%.