Authors: Zeyu Chen; Xiangfeng Luo; Yan Sun
Addresses: School of Computer Engineering and Science, Shanghai University, Shanghai, China ' School of Computer Engineering and Science, Shanghai University, Shanghai, China ' School of Computer Engineering and Science, Shanghai University, Shanghai, China
Abstract: The performance of deep neural network for object detection depends on the amount of data. In the field of maritime object detection, the diversity of weather, target scale, position and orientation make real data acquisition hard and expensive. Recently, generating synthetic data is a new trend to enrich the training set. However, synthetic data might not improve the detection accuracy. Two problems remain unsolved: 1) what kind of data need to be augmented; 2) how to augment synthetic data. In this paper, we utilise knowledge-based rules to seek effective synthetic samples. Herein, we propose two synthetic data augmentation rules: 1) what to augment depends on the gap between training and expiring data distribution; 2) the robustness and effectiveness of synthetic data depends on the proper proportion and domain randomisation. The experiments show that the average accuracy of boat classification increases 3% with our synthetic data in Pascal VOC test set.
Keywords: data augmentation; synthetic data; synthetic data augmentation rules; virtual environment; object detection; domain randomisation.
International Journal of Computational Science and Engineering, 2020 Vol.23 No.2, pp.169 - 176
Received: 31 Jan 2020
Accepted: 05 Mar 2020
Published online: 23 Oct 2020 *