Improving Object Detection in VHR Aerial Orthomosaics

Abstract

In this paper we investigate how to improve object detection on very high resolution orthomosaics. For this, we present a new detection model ResnetYolo, with a Resnet50 backbone and selectable detection heads. Furthermore, we propose two novel techniques to post-process the object detection results: a neighbour based patch NMS algorithm and an IoA based filtering technique. Finally, we fuse color and depth data in order to further increase the results of our deep learning model. We test these improvements on two distinct, challenging use cases: solar panel and swimming pool detection. The images are very high resolution color and elevation orthomosaics, taken from plane photography. Our final models reach an average precision of 78.5% and 44.4% respectively, outperforming the baseline models by over 15% AP.