Increased Robustness of Object Detection on Aerial Image Datasets using Simulated Imagery

Machine learning-based models for object detection rely on large datasets of labeled images, such as COCO or ImageNet. When models trained on these datasets are applied to aerial images recorded on Unmanned Aerial Vehicles (UAVs), the problem arises that the conditions under which the training images were created (for example, light, altitude, or angle) may be different in the environment where the UAVs are put into practice, leading to failed detections. This problem becomes even more pressing in safety critical applications where failures can have huge negative impacts and also constitutes an obstacle for certification of cognitive components in UAVs. Along a case study on car detection in low-altitude aerial imagery, we show that using, both, artificial and real images for model training has a positive effect on the performance of object detection algorithms when the trained model is applied on images from another domain. Since simulated images are easy to create and object labels are inherently given, the presented approach shows a promising direction for scenarios where adequate datasets are difficult to obtain, as well as for targeted exploration of weak points of object detection algorithms.