Synthetic data of aerial images of crocodiles

Image object detectors are a powerful tool that allows researchers and farmers to automatically detect and analyse animals in aerial images using unmanned aerial vehicles. This technology reduces the cost and time required to do animal surveys. However, one major challenge of object detectors is the large amount of data required to train the network. This can especially become problematic if the target animal is scarce and there is no available data to train on. In this paper, the generation and use of synthetic data for training a you only look once (YOLO) object detector to identify crocodiles in aerial images were explored. Additional methods for improving the performance of the detector were also presented. The results demonstrated that the accuracy of all the models was improved when the backbone layers were frozen during training, except when training on a large real-world dataset. Training on only synthetic images is possible, but training on a small amount of real-world images outperformed the synthetic data. Fine-tuning after initially training on a synthetic dataset, further improves the model accuracy even if it is a small number of real-world images, and produces a model better than if the model only trains on that small number of real-world images. Even better accuracy was obtained when training on a mixed dataset that includes synthetic as well as a small number of real-world images. Synthetic data can thus significantly improve object detectors when working with sparse data, and this approach was effective for an object detection system tasked with detecting crocodiles in aerial images.