Thesis Projects
Applications for Bachelor and Master theses are open.
How to Apply
If you are interested, please send your application to laehner at uni-bonn.de. The application should include
- a short introduction of yourself, your study program and semester, why you are interested in a thesis with us, programming languages you know
- which project your are interested in, or a short proposal of your own topic (with an explanation why you think it should be supervised by us)
- your transcript of records
- optional: a CV
Available Projects
Hierarchical Geometric Pooling for 3D Mesh GNNs
Treating 3D meshes as graph data within geometric deep learning model has emerged as a common practice in recent years. However, multiple contributions such as MeshCNN showed that specializing the architecture to the sub-domain (surfaces rather than general graphs) improves the model’s performance. This project is focused on defining geometric pooling operators similar to those of 2D image data, via the hierarchical application of mesh simplification algorithms. Adequate programming skills in C++ are recommended.
Contact: Daniele Baieri (dbaieri at uni-bonn.de)
3D Reconstruction with Material Priors
Previous 3D reconstruction methods for indoor scenes from RGB images typically do not incorporate any prior knowledge about the materials present in the scene. Meanwhile, there are existing methods that can predict the material of a scene directly from its image. This thesis explores the impact of initializing the material using predictions from a generative model, rather than starting from scratch. Our hypothesis is that this prior knowledge can reduce the inherent ambiguity in the reconstruction problem and improve the overall reconstruction quality.
Contact: Mohamed Ebbed (mebbed at uni-bonn.de)
Segmentation on the Go for 3D Gaussian Splatting
Reconstruction and novel view synthesis with 3D Gaussian Splatting often suffers in the presence of occlusion and less sharp details in the foreground objects. This thesis will explore how segmentation and virtual cameras (generating new eg. deblurred training images) can improve results in complex scenes.
Contact: Hongyu Zhou (hzhou at uni-bonn.de)
Geometry opacity meets SDF in Gaussian Splatting
Geometry opacity is found to improve the representation power for Gaussian Splatting especially for transparent objects. However, it is not widely accepted and used yet. This thesis aims to extend geometry opacity to more geometry-aware Gaussian Splatting varieties to test its effectiveness.
Contact: Hongyu Zhou (hzhou at uni-bonn.de)
Spectral Rewiring for Graph Neural Networks
Graph Neural Networks (GNNs) learn by repeatedly exchanging information between neighboring nodes. While highly successful, deep GNNs are often limited by oversquashing, where information from distant parts of a graph is forced through a single edge, and oversmoothing, where node representations become indistinguishable.
Recent work has shown that modifying the graph structure can substantially improve information flow. In this project, we investigate spectral graph rewiring approaches that learn new graph connections by smoothly deforming the graph spectrum. The goal is to develop principled rewiring strategies that preserve geometric structure while reducing bottlenecks and improving message passing.
The project involves studying spectral graph theory, graph curvature, and modern GNN architectures, as well as implementing and evaluating rewiring methods on real-world graph learning tasks.
Contact: Ipsit Mantri (kmantri at uni-bonn.de)