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

Abstract

Machine learning-based models for object detectionrely on large datasets of labeled images, such as COCO orImageNet. When models trained on these datasets are appliedto aerial images recorded on Unmanned Aerial Vehicles (UAVs),the problem arises that the conditions under which the trainingimages were created (for example, light, altitude, or angle) maybe different in the environment where the UAVs are put intopractice, leading to failed detections. This problem becomes evenmore pressing in safety critical applications where failures canhave huge negative impacts and also constitutes an obstaclefor certification of cognitive components in UAVs. Along a casestudy on car detection in low-altitude aerial imagery, we showthat using, both, artificial and real images for model traininghas a positive effect on the performance of object detectionalgorithms when the trained model is applied on images fromanother domain. Since simulated images are easy to create andobject labels are inherently given, the presented approach showsa promising direction for scenarios where adequate datasets aredifficult to obtain, as well as for targeted exploration of weakpoints of object detection algorithms.