This post-doc aims at reconstructing the 3D earth geometry from satellite imagery by dense matching. If this technique is well developped now, it still fails to produce a good quality reconstruction on certain difficult cases such as strongly repetitive textures, textureless areas, specular surfaces and on strong depth discontinuities which are often encountered in urban areas and on the forest canopy. Recent works have demonstrated that the use of Deep Learning and in particular Convolutional Neural Networks (CNNs) techniques allow for a significant improvement of dense matching in such difficult cases. Thus the aim of this post-doc is to leverage on such techniques in order to increase the quality of the reconstruction and to handle these issues better.
The input data for the post doc will consist of oriented aerial or satellite images over urban and forested areas as well as LiDAR scans of the same areas to serve as ground truth and learning data. The main objective is to exploit this learning data in order to develop a dense matching method specifically adapted to handling vertical images of forest.
All information on this proposal are available at:
(c) GdR 720 ISIS - CNRS - 2011-2019.