Student of Natural Sciences, Engineering or Computer Science (f/m/d)

The Institute of Remote Sensing Methodology is a DLR institute with sites in Oberpfaffenhofen near Munich, Berlin-Adlershof and Neustrelitz in Mecklenburg-Western Pomerania. Together with the German Remote Sensing Data Centre, the Institute forms the Earth Observation Centre (EOC), Germany’s centre of excellence for Earth observation.

What To Expect

This research proposes a diffusion-based generative framework for reconstructing building roof planes directly from very high–resolution (VHR) true-ortho aerial imagery. Inspired by the architectural layout synthesis paradigm of HouseDiffusion, the model instead predicts structured roof-plane configurations by conditioning the diffusion process on both image features and graph-based spatial relationships between neighboring roof elements. A roof adjacency graph encodes geometric and topological constraints (e.g., shared ridges, plane orientation continuity), enabling the model to incorporate contextual structural priors during generation. The diffusion process iteratively refines candidate roof plane representations while respecting these relational constraints. This formulation enables structured roof reconstruction from purely 2D observations without explicit 3D supervision, bridging generative modeling with geometric reasoning in remote sensing.

Your task

• Masterthesis for Diffusion-based roof plane generation from very-high resolution imagery

Your profile

• Ongoing Master’s studies in Computer Science, AI, Computer Vision, Geoinformatics, or a related discipline
• Programming experience in Python
• Experience in AI and Deep Learning approaches
• Experience with Pytorch or Tensorflow frameworks
• Good communication skills and proficiency in English (spoken and written)
• Basic knowledge of semantic segmentation and object detection is a plus