Master student @ uulm
LangueViz is a system developed in a Computer Vision course to detect spoken language from audio using mel spectrograms and a DenseNet-based CNN. The project aimed to classify audio clips visually rather than using traditional methods. Despite achieving good validation accuracy, the model struggled with real-world inputs, revealing challenges with dataset bias and generalization.
ParkDas is a Telegram-based parking assistant developed for the University of Ulm campus as part of a Multimodal UIs course. It combines natural language processing, location-aware recommendations, and map-based outputs to help users find parking quickly and conveniently. The project emphasizes usability through conversational interaction and visual feedback, enhancing user experience while navigating campus parking.
Bar-JEPA is a method developed for extracting numerical values from bar charts using self-supervised learning to enhance chart understanding despite a lack of labeled data. By employing finetuned I-JEPA encoders and synthetic data, the approach effectively improves value recovery, demonstrating the importance of high-quality representations in machine learning tasks.
During my master’s program, I enhanced my bachelor’s thesis on photorealistic rendering by implementing light source estimation. This research improved realism by addressing issues like static lighting and geometry clipping. Using a combination of Python tools, I successfully estimated directional and point lights, resulting in more coherent and believable scenes.
The thesis titled "Photorealistic Rendering of Training Data for Object Detection and Pose Estimation" discusses developing a data generation pipeline that integrates real and synthetic imagery using a physics engine. It highlights challenges in pose estimation and presents results indicating the generated data's potential for training detection models, despite certain limitations.
A cyberpunk 2D platformer featuring a modding system, developed for a game jam.
In a 3D modeling and interactive visualization project, a team developed ARchitecture, an augmented reality application for architectural visualization using Unity AR Foundation. The app allows users to explore BIM data interactively, featuring dynamic models and extensive real-time rendering techniques, significantly improving the presentation of architectural information over traditional methods.
A local asymmetrical multiplayer game inspired by "Prop Hunt", developed for a game jam.
A collection of physics simulations, developed for a university class.
A full-body Kinect game that was created for two university classes, built using only custom shaders.
An outline of contributions to ATX's NH Software, which aids in safely testing circuit boards. The software helps estimate project complexity and enables precise NH placement, reducing human error. Key contributions include enhancing shape support, automating outline extraction, and overhauling the rendering engine for improved performance and responsiveness.
A game developed during "Ludum Dare 42" over the course of a weekend.
A location-based mobile game for Android, featuring interesting social features and gameplay.
A small game engine, that involved developing a first-person shooter game with a custom C++/DirectX 11 implementation. The project comprised a terrain generator and gameplay systems, utilizing techniques like diamond-square for height maps. It emphasized real-time rendering, lighting, and asset handling, providing a solid foundation in graphics programming and engine development.
A VR game created during a "Semester Games Jam" over the course of a weekend.
A mobile game for iOS and Android developed for a university class, featuring a new take on the game of dominos.
A game developed during a "Semester Games Jam" over the course of a weekend.
The Lomecs project at ATX Hardware GmbH aimed to develop a custom machine for testing circuit boards and capturing data. Challenges included integrating hardware and software, ensuring precise motion control, and automating measurements. Ultimately, the focus shifted to successful optomechanical imaging after measurement precision proved inadequate. The project provided valuable engineering experience.
