Learning from synthetic data for object detection on aerial images

Camera-equipped Unmanned Aerial Vehicles (UAVs) have been developed as autonomous vision systems and are widely used in various fields such as surveillance, search and rescue, and agriculture. Object detectors need to be robust to changing light, weather, and different application scenarios in real-time applications. Object detection algorithms have evolved considerably in these years, however the performance of object detection for UAVs has lagged behind. As it is largely limited by the dataset. Obtaining manually labeled object detection datasets on UAVs is undoubtedly expensive, and requires comprehensive consideration of the UAV's position, the object's size, and the weather, etc. Meanwhile, UAVs are prohibited from flying in certain scenarios, and the collection of data also raises privacy issues. To address the above problems, this paper proposes a method to enhance the data based on the Unreal Engine (UE) and Airsim synthetic dataset. We also take a Generative Adversarial Network (GAN)-based domain adaptive approach to make reduce the domain difference between synthetic and real data.