ECopy-Paste: An Effective Data Augmentation Method for Object Detection of UAV
Lili Zhang, Xiulei Liu, Qiang Tong · 2022 International Conference on Big Data, Information and Computer Network (BDICN) · 2022
Deep learning-based target models require a large amount of data for training in order to achieve good generalization. But the amount of data in some widely used datasets cannot meet this requirement, while the data set also has the problems of brightness imbalance, small targets and data imbalance. These will lead to poor robustness of the model. In this paper, we propose the ECopy-Paste method to solve the above problems. Our method consists of three parts: brightness transformation, over large scale jittering and copy-paste. Brightness transformation can balance the brightness of datasets, Oversize large scale jittering can expand the size of small targets and Copy-paste can relieve the problem of data imbalance and enlarge the size of datasets. Our method tested on VisDrone2019-DET by using Faster R-CNN, and the final score obtained is 17.6%, which is about 3.1% improvement over the benchmark of VisDrone2019-DET.