Joint Distribution and Class-based Data Augmentation for Wildlife Detection

Data augmentation is of great importance to alleviate the insufficiency of training samples, and further improve wildlife detection accuracy. However, current data augmentation methods tend to augment all kinds of samples equally, ignoring the problem of uneven distribution of the number and size of all kinds of samples in wildlife detection datasets, resulting in poor generalization of the model. To address this problem, this paper proposes a joint distribution and class-based data augmentation method for wildlife detection. In this method, diverse rather than universal data augmentation methods are introduced for different classes with a small proportion. This makes the distributions of different classes more balanced. Therefore, each class even with a small number of samples gets good training as well. To evaluate the effectiveness of the proposed method, extensive comparative experiments are conducted. Experimental results show the superiority of our proposed method. Specifically, the detection accuracy of Faster RCNN with Swin Transformer as the backbone network is improved by 0.8% to 96.2% after data augmentation with our method.