Vansteelandt is a geo-surveying company which flies over Flanders on a yearly basis, capturing high resolution aerial imagery. Looking for a way to further exploit this data, they started a project with the EAVISE research group in order to develop AI algorithms, capable of extracting various metadata from the images.
Building on my previous experience with satellite imagery, I firstly developed a pipeline to be able to train and run deep learning models on these aerial images, automatically handling the division of the images in smaller the chunks, whilst keeping the possibility to transform the output of the models to proper geographical coordinates. This enabled various researchers to develop a plethora of AI models, capable of extracting different metadata from the aerial footage. Finally, I used this pipeline to research the potential of object detection on this high resolution data. More specifically, I developed and trained convolutional neural networks capable of detecting solar panels and swimming pools. Furthermore, I designed a methodology to fuse RGB and depth data, increasing the accuracy of the detection models.
References
Improving Object Detection in VHR Aerial Orthomosaics
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.
@inproceedings{ophoff_plane,title={Improving Object Detection in VHR Aerial Orthomosaics},author={Ophoff, Tanguy and Van Beeck, Kristof and Goedem{\'e}, Toon},booktitle={Proceedings of the European Conference on Computer Vision Workshops (ECCV Workshops)},year={2023},editor={Karlinsky, Leonid and Michaeli, Tomer and Nishino, Ko},publisher={Springer Nature Switzerland},isbn={978-3-031-25066-8},}