Teaching

Interactive Computational Fluid Dynamics and Rendering

[Type: Bachelor or Master Thesis, Focus: Simulation, Computer Graphics, Machine Learning]

​

Over the course of the last year, we developed a framework that is capable of computing and rendering fluid simulations at interactive rates on the GPU (Rust/WGPU/WGSL). The framework combines fluid dynamics with thermodynamical and chemical processes to simulate combustion and even fire extinguishing. We already published a paper at SIGGRAPH Asia based on this work (link). We enjoy working together with students to implement related concepts and prototype new ideas. Current areas of interest include, but are not limited to, the combination of a classical simulation with machine learning approaches, the implementation of state-of-the-art simulation concepts and real-time capable rendering techniques for volumetric rendering. Possible topics include concepts like: PINN, multi-grid pressure projection, voxel splatting, GPU based sorting, radiative heating, GPU based bounding volume hierarchies. We also encourage you to propose related concepts that you would be interested in.

​

Automatic Robot-driven Map Generation

[Type: Master Thesis, Focus: Robotics, Machine Learning, Computer Vision]

​

This project aims to develop a system for automatic 3D indoor scene reconstruction using a quadruped robot equipped with onboard cameras. The robot will either autonomously or manually explore indoor environments and build compact 3D models, similar to "dollhouse views," using modern techniques such as Gaussian Splatting or Neural Radiance Fields. Key challenges include robust SLAM, trajectory planning, and real-time scene representation from egocentric visual data. The resulting system enables applications in inspection, digital twins, or VR/AR without requiring external scanning hardware.

​

Gesture-based Robot Control

[Type: Bachelor or Master Thesis, Focus: Robotics, Machine Learning, Computer Vision]

​

This project explores intuitive human-robot interaction through vision-based gesture recognition. A camera-equipped robot will be trained to recognize a set of predefined hand/body gestures in real time and map them to locomotion or manipulation commands. The system will use neural networks or transformer-based models for gesture detection, trained on either public datasets or custom recordings. The goal is to enable hands-free control for tasks like guiding, commanding, or stopping the robot, particularly useful in assistive or social robotics contexts.

​

​Proprioceptive Touch Localization for Legged Robots Using Data-Driven Methods

[Type: Bachelor or Master Thesis, Focus: Robotics, Machine Learning]

​

This project aims to develop a data-driven whole-body touch sensing system for a quadruped robot using only proprioceptive signals, without requiring additional sensors. Inspired by the UniTac approach, the system will be trained on real-world data collected by manually inducing contact at known body locations and recording joint torques and positions. A neural network will then be trained to predict the location of contact from this proprioceptive input, enabling accurate touch localization and enhancing physical human-robot interaction without modifying the robot’s hardware.

​

Neural Marching Cubes

[Type: Bachelor or Master Thesis, Focus: Machine Learning]

​

The objective of this project is to investigate the potential of combining implicit neural representations with traditional surface extraction techniques, specifically through the use of Neural Marching Cubes, for high-fidelity 3D shape reconstruction. Neural Marching Cubes leverages a neural network to learn a continuous signed distance function (SDF) or occupancy field that defines a 3D surface implicitly in space. In this project, we explore how a neural network can be trained to approximate the underlying geometry of a 3D object from sparse or noisy data such as point clouds or multi-view images. Once trained, this continuous representation allows for the application of the classic Marching Cubes algorithm on a dense sampling of the learned field, producing a mesh that accurately captures the shape's fine-grained details.This approach bridges deep learning with classical computer graphics, enabling high-resolution 3D surface reconstruction with improved generalization and topological consistency. Ultimately, the project aims to demonstrate the effectiveness of neural fields in 3D modeling workflows, and how their integration with structured extraction techniques can advance the state-of-the-art in 3D reconstruction and generation.

​

Image Generation Using GPT-Based Model and VQ-VAE

[Type: Bachelor Thesis, Focus: Machine Learning]

​

The objective of this project is to investigate the potential of combining a transformer-based model, such as GPT, with a Vector Quantized Variational Autoencoder (VQ-VAE) for the purpose of image generation. VQ-VAE is a type of generative model that effectively compresses images into discrete latent representations, making them suitable for sequence modeling.In this project, we explore how the discrete latent codes produced by a trained VQ-VAE can be treated similarly to tokens in a language model. By using a GPT-style transformer to model the distribution over these latent codes, we aim to generate new images in a sequential, token-by-token manner. This approach leverages the powerful generative capabilities of transformers in the domain of visual data, allowing for the creation of coherent and high-quality images. Ultimately, the project aims to provide insights into the synergy between VQ-VAE and transformer-based models, demonstrating how they can be integrated to push the boundaries of generative image modeling.

​

Monocular Fire Reconstruction

[Type: Master Thesis, Focus: Computer Vision, Machine Learning]

​

Traditional reconstructions of fire are based on a set of inward-facing cameras. The large baselines enable precise triangulation of the individual flames. This, however, comes with the main disadvantage of cost and intense calibration. For this reason, this project aims to explore the possibility of significantly reducing the amount of required cameras by introducing powerful pre-trained methods, such as monocular depth, point tracking, and diffusion-based novel-view-synthesis. While approaching the target of using only a single camera, photometric losses as well the generalization capabilities of the reconstructed model will be evaluated.

 

Prediction of Bone Healing Processes using Sparse Longitudinal Data

[Type: Master Thesis, Focus: Computer Vision, Medical, Machine Learning]

​

The aim of this project is to investigate whether bone healing processes can be predicted on the basis of sparse longitudinal data, both in terms of overall extent and temporal resolution. To this end, a dummy CT data set must first be created, on which the various AI models will then be trained and tested.

​

Bone Strength Analysis (FemoraLyze)

[Type: Bachelor or Master Thesis, Focus: Computer Vision, Medical, Machine Learning]

​

FemoraLyze is a modular framework for a fully-automated proximal femur analysis. Based on different segmentation masks, the framework is able to collect metrics such as cortical thickness in different bone areas, volumes, angles and distances as well as bone structure parameters. These are relevant in medical prognosis, diagnosis and therapy. In order to further increase the benefits for medical professionals, this framework is to be expanded with additional modules (e.g., finite element analysis) and modalities (e.g., MRI or clinical CT). Runtime optimization and a GUI would also be desirable.

​

AI-based Implant Modeling

[Type: Bachelor or Master Thesis, Focus: Computer Vision, Medical, Machine Learning]

​

The Institute for Digital Implant Research (IDIR) is investigating how digital twins can be used to improve the development of patient-specific implants. This includes the sub-projects aneurysms, bone nails and heart valves. They all have the same approach in common: We aim to create a digital twin from real data (e.g., CT, MRI) of the patient. The optimum implant for the patient must now be determined/developed by simulation, which then only needs to be manufactured and implanted in the final step. In this context, there are always exciting questions that could be dealt with in the course of a Bachelor's or Master's thesis. If you are interested, please get in touch and we will see if we can find a suitable topic.

​

Learning based reverse engineering of 3D file formats

[Type: Bachelor or Master Thesis, Focus: Machine Learning]

​

This project aims to develop a neural-network architecture that receives a binary 3D file of unknown format as input and outputs a file format specification. To train this network, a method to create synthetic data should be developed. On one hand, arbitrary file format specifications can be synthesized and on the other hand 3D files can either be created procedurally or by using a dataset of 3D objects which are then augmented and converted to the previously synthesized file format specifications.​

​​

​​​Uncertainty Visualization for AI-based Segmentations of Medical CT Volumes

[Type: Bachelor or Master Thesis, Focus: Computer Vision, Medical, Machine Learning]

​

Various organs of the human patient can be segmented from CT scans and the corresponding greyscale axial views. AI models already achieve a high accuracy in the segmentation, but still need to be corrected. Instead of a binary classification mask as output, a heatmap may display the certainty of the segmentation and improve the understandability of the model.

​​

Effect of the Patient's Rotation Angle on Tissue Deformation in CT Volumes

[Type: Bachelor or Master Thesis, Focus: Computer Vision, Medical]

​

While medical CT scans are generally taken in a supine position (lying on the back), during the actual surgery, patients are often positioned on their side. This leads to a deformation of internal organs so that their position does not match the CT scan anymore. The elasticity of various organ types may be estimated from tissue characteristics and used to predict absolute movement under the influence of gravity.

​

Semantic Segmentation of Outdoor Scenes for Robot Navigation

[Type: Bachelor Thesis, Focus: Computer Vision]

​

This project aims to develop a deep learning model for semantic segmentation in outdoor environments, with a specific focus on identifying and classifying various types of vegetation such as trees, grass, shrubs, and other plant species. The student will work with datasets of outdoor images and leverage convolutional neural networks (CNNs) or other machine learning models to differentiate vegetation from non-vegetation areas. The outcome of the project will be a tool that can accurately map vegetation in outdoor scenes, potentially useful for applications like environmental monitoring, agriculture, and urban planning.

​

​​Randomized Look-Up Data Structures for Spatial Data

[Type: Bachelor or Master Thesis, Focus: Computer Graphics]

​

The goal of this project is to investigate how randomization can be used to engineer efficient data structures for queries on spatial data. Initially, a literature research should be conducted to compile an overview of what the state of the art is. Based on this, data structures that have only been proposed theoretically should be implemented or entirely new data structures should be developed and implemented. Experiments should then be conducted comparing the performance to that of conventional data structures (e. g. Kd-trees, octrees). Optionally, these newly developed data structures could make use of learned biases to make random decisions which might yield further performance improvements.​