Data Augmentation Based on Cannikin Law for Object Detection
Xue Feng Deng, Hongwei Luo, Hao Zhao, Jing Zhang, Shan Yang, Boping Mei, Hua Zhang · 2021
Data augmentation is one of the most significant approaches to improve the performance of object detection. Currently, data augmentation is generally used for resolving the long tail and sample imbalance problem. However, there is an imbalance of detection performance in multi-class object detection without the impact of the above-mentioned problems. In this paper, we present an offline data augmentation method based on Cannikin's Law for augmenting the categories far below the average detection accuracy. Specifically, we first apply the statistical method to find the categories with low detection performance. Secondly, a data augmentation following the cut-paste principle is constructed. The instances from the specific categories are segmented and placed into the selected pictures in the training dataset. Lastly, the augmented pictures are selected by the similarity measurement mechanism. Extensive experiments are conducted on the Pascal VOC dataset. For the FCOS detector, the mean Average Precision (mAP) is increased from 79.12% to 83.14%. For the RetinaNet detector, the mAP is boosted from 81.6% to 83.57%. Meanwhile, the detection accuracy of the specific categories is improved by 10% at least. The experimental results show the effectiveness of our data augmentation method.