Training Object Detectors with Synthetic Data for Autonomous UAV Sampling Applications

Abstract

Training accurate object detection models often requires a large amount training data. In some cases, limited imagery, from drastically different perspectives than the desired target view positions and angles, may be available for specific objects of interest. Training accurate models with this imagery may not be possible and require a lot of performance-limiting assumptions. However, it may be possible to use this limited imagery to create a 3D model of the targets and their surrounding area. In this paper, we explore training an object detector using only synthetic imagery to detect rooftop stacks for UAV sampling tasks. We show that this detector performs well on real imagery, and enables autonomous UAV sampling. We also note that this approach is general, and should extend to other objects.