Reconstruction automatique de contours de toits en 3D à l’aide du Frame Field Learning

Abstract

Several AI models, such as Frame Field Learning, have been developed recently to detect buildings' outlines on aerial images. However, they are not able to predict a height estimation for these buildings. In addition, due to crossfall, a planimetric error between the detected position on the aerial image and the actual position may exist. Moreover, IGN owns oriented images which are images acquired from aircraft. Their primary purpose is to be used in the production of BD ORTHO®. Their interest is that each building is seen several times from different viewpoints. This multiplicity of viewpoints enables 3D reconstruction. We propose an algorithm which uses Frame Field Learning or Pix2Poly inferences outputs from oriented images to extract building outlines in 3D. The code is available here : https://github.com/IGNF/Samon_gouttieres