• Ingenieurpraxis | Engineering Practice

    (info) Information about writing a engineering practice can be found here

    Archive

  • Bachelorsthesis

    (info) Information about writing a bachelorsthesis can be found here.

    Archive

    • Local Public Transport Server

      Student

      Vorname: Andreas

      Name: Schrägle

    • Teleoperated control of a humanoid robot

      Student/in

      Vorname: Jens

      Name: Ernstberger

    • Poppy Kinect

      Student/in

      Vorname: Max

      Name: Musterstudi

    • autoMAP

      Student/in

      Vorname: Simon

      Name: Schumann

    • Scene Detection in Skiing Videos

      Student/in

      Vorname: Niklas

      Name: Walther

    • 3D Print Augmented Reality

      Student/in

      Vorname: Max

      Name: Musterstudi

    • Real-time LRV speed estimation from sigle forward facing camera

      Student/in

      Vorname: Steffen

      Name: Büttner

    • Anitblockier Bilder

      Student/in

      Vorname: Niklas

      Name: Schwarz

    • AirHockey Unity Simulation

      Student/in

      Vorname: Max

      Name: Musterstudi

    • Object Classification Autonomous Vehicle

      Student/in

      Vorname: Max

      Name: Musterstudi

    • Adaptive Hintergrundbeleuchtung für Monitore/Fernseher

      Student/in

      Vorname: Diana

      Name: Leichte

    • Auralisation von Bewegung

      Student/in

      Vorname: Flora

      Name: Latz

    • Segmentation of Medical Images with Deep Learning

      Student/in

      Vorname: Vincent

      Name: von Büren

    • GANs learning to draw oil paintings

      Student/in

      Vorname: Beyrem

      Name: Kaddech

    • Optimierung der Architektur neuronaler Netze für modulare Probleme

      Student/in

      Vorname: Anna

      Name: Schua

    • Analysis Pipeline for Quantitavie Mapping of Physical Parameters in Glioma Patients

      Student/in

      Vorname: Jim

      Name: Zhao

    • Vorsortierung von Fotos

      Allgemeine Daten

      Student

      Vorname: Lorenz

    • Measuring Accurate Timings in Outdoor Sports Using Photoelectric Sensors and Ultrasound

      Student/in

      Vorname: Tim

      Name: Spickermann

    • Audio Analysis and Transcription

      Student/in

      Vorname: Lisa

      Name: Gan

    • Extreme Learning Machines with Structured Matrices

      Student/in

      Vorname: Ruslan

      Name: Mammadov

    • Automatisierung eines 3D-Druck-Labors

      Student/in

      Vorname: Matthias

      Name: Emde

    • Application for the Analysis of Portrait Composition

      Student/in

      Vorname: Marisa

      Name: Ripoll

    • Ridge Regression for Big Data Applications: Trade-Off between Memory Consumption and Learning Speed

      Student/in

      Vorname: Alexander

      Name: Schindler

    • Formative Evaluation of Data Management Tools on Data Quality and Usability

      Student

      First Name: Hian Zing

      Last Name: Voon

    • Sequential Context-Aware Action Recognition on Film

      Student

      Unknown User (ge78diq) 

      Email: kerim.anater@tum.de

    • Comparison of 3D Models (Jakob Schödl)
      Student:Jakob Schödl

      Abstract:

      In this thesis, I explore different methods for the comparison of 3D models, with the aim of benchmarking the capabilities of Large Language Models (LLMs) in generating 3D shapes. OpenSCAD, a descriptive functional language for CAD models, was used as the representation of the compared and generated shapes.

      I decided to implement three comparison approaches: Firstly, the comparison of a set of global features of the triangle mesh, such as its volume or bounding box ratios; secondly, a comparison based on the geometric intersection of both 3D models; and thirdly, an analysis of the topological structure using Multiresolutional Reeb Graphs (MRGs).

      For evaluating the approaches, six types of perturbations were applied to numerous 3D shapes. The distance of the modified shapes to their unmodified version was used to characterise their reaction to different possible faults in LLM-generated 3D models, such as missing parts or invalid proportions.

      Each of the methods capture different aspects of similarity. While the volumetric distance yields results that are easily predictable and explainable, the comparison of global features is the computationally most lightweight solution. Comparing MRGs correlates with the human perception of similarity the most, as the topological structure of an object is taken into account.

      This work provides a foundation for benchmarking and fine-tuning any LLMs that are used for generating 3D models.

      Email:jakob.schoedl@tum.de
      Status:

      STARTING

      Supervisor:

      David Fresacher 

    • Better Data, Better Models: A Quality Assessment Pipeline for Code-Based CAD Data
      Student:Malin Wen Xia

      Abstract: 

      High-quality datasets are essential for reliable and trustworthy AI systems. In the context of AI-driven 3D model generation, OpenSCAD - a script-based parametric CAD language - offers a unique opportunity due to its textual and modular structure, but its publicly available datasets often suffer from low quality, redundancy, and legal ambiguities. This thesis investigates how OpenSCAD datasets can be systematically assessed and curated to improve their suitability for AI training.

      A comprehensive quality assessment pipeline was developed and applied to 72,197 OpenSCAD projects (containing 189,158 .scad files in total) sourced from GitHub, Printables, and Thingiverse. The pipeline integrates legal, technical, and semantic evaluation stages, including license filtering and PII detection, dependency resolution and compilation checks, semantic and heuristic analysis, deduplication, and comment-based labeling. Each step is logged and documented to ensure reproducibility.

      After processing, the dataset was reduced to 55,573 high-quality projects with one .scad file each, significantly improving reliability, relevance, semantic richness, legal compliance, and interpretability, while maintaining design diversity. Compilation and dependency checks guaranteed syntactic validity, semantic filters removed low-information or mesh-like files, and deduplication minimized redundancy. NLP-assisted labeling and comment analysis enhanced readability and usability for downstream AI applications.

      The results demonstrate that domain-specific, automated quality assessment transforms raw OpenSCAD collections into curated, AI-ready datasets, operationalizing Trustworthy AI principles in practice. This work provides both a replicable methodology and a conceptual framework for quality assurance in code-based 3D design data, laying the foundation for downstream model evaluation and adaptation to other CAD languages or 3D formats.

      Email:malin.xia@tum.de
      Status:

      STARTING

      Supervisor:

      David Fresacher 

    • Test 1 2 3
    • Benchmarking Spatial Reasoning of LLMs in Text-to-3D Modeling
      Student: Barış Can Sönal 

      Abstract: Large Language Models (LLMs) have achieved significant success in software engineering. However, their capacity for 3D spatial reasoning and procedural modeling remains underexplored. While current text-to-3D methods are effective at generating visual representations such as meshes or point clouds, these outputs usually lack parametric editability and underlying construction logic. As a result, generating procedural Computer-Aided Design (CAD) models directly from text remains a complex and largely overlooked problem. This thesis evaluates the 3D spatial reasoning capabilities of open-weight LLMs in order to address this challenge.

      A fully automated benchmarking framework was created to measure the generation of procedural OpenSCAD geometry based only on natural language descriptions. The evaluation uses a balanced dataset of 900 3D designs with reference source code and structural metadata for different complexity levels. Before assessing geometric performance, a two-stage spatial alignment process is applied. This process uses Principal Component Analysis (PCA) and Iterative Closest Point (ICP) registration to evaluate geometries on a coordinate-independent basis.

      Furthermore, a new Composite Success Score ($S_c$) is introduced. It combines volumetric overlap, surface similarity, and syntactical compilation validity into one metric. By comparing state-of-the-art models, including Qwen2.5-Coder:32B, Codestral:22B, and DeepSeek-Coder-V2-Lite, using Zero-Shot and Chain-of-Thought prompting strategies, this research proposes a robust and scalable methodology for evaluating true spatial reasoning in generative AI.
      Email: baris.soenal@tum.de
      Status:

      COMPLETED
      Supervisor: David Fresacher 
  • Forschungspraxis | Research Internship

    (warning) Detailed information can be found here

    Objective

    The research internship is intended to prepare students for their later scientific activities in research and development by providing them with insight into current research topics during their master’s studies. The research internship takes the form of project work. Each student in the master’s program works on an individual project assigned by a professor/supervisor from the respective Chair (cf. Procedure) within the scope of the research internship. Upon successful completion of the internship students will be able to draw up (apply learned concepts) and plan a project with an engineering character, to define milestones, as well as to document the progress and results of the project and to present them to an audience.

  • Mastersthesis

    (info) Information about writing a mastersthesis can be found here.

    Archive

    • Sequential Recommender Systems

      Student/in

      Vorname: Tim

      Name: Bicker

    • Modification and Testing of a SLAM Framework for Dynamic Environments

      Wiedemann

      Heike 

      Email: heike.wiedemann@tum.de

    • Multi-Agent Deep Reinforcement Learning

      Student/in

      Vorname: Sven

      Name: Gronauer

    • Correlation between physiological data and video features to better understand video induced emotions

      Student/in

      Vorname: Carla

      Name: Pregel Hoderlein

    • Improvement of Deep-learned Object Detectors

      Student

      Vorname: Siyuan

      Nachname: Liu

    • Self Supervised Learning for Emotion Prediction induced by Videos

      Student/in

      Vorname: Alexander

      Name: Lechthaler

    • Machine Learning for Time Series Prediction in the Financial Industry

      Student/in

      Vorname: Arturo Buitrago

      Name: Méndez

    • On Generating Mathematical Formulae

      Student/in

      Vorname: Luca

      Name: Sacchetto

    • Detection of Teeth Grinding and Clenching using Surface Electromyography

      Student/in

      Vorname: Hella

      Name: Toto-Kiesa

    • Learned Under Water Feature Matcher based on Generated Artificial Image Base

      Student/in

      Vorname: Patricia

      Name: Schöntag

    • Latency Prediction for Wireless System Synchronisation

      Student/in

      Vorname: Julian

      Name: Eder

    • Deep Learning for Financial Time Series Prediction

      Student/in

      Vorname: Mirko

      Name: Corletto

    • On the Assessment of RL

      Student/in

      Vorname: Tobias

      Name: Allgeier

    • Federated Learning in Pedestrian Trajectory Prediction Tasks

      Student/in

      Vorname: Claas

      Name: Brüß

    • Distilling Neural Networks for Real-Time Drone Control

      Student/in

      Vorname: Anna

      Name: Schua

    • Causal Regularization in Deep Learning Using the Average Causal Effect
      Student:Kathrin Khadra

      Abstract: Causal Interpretability aims to make decisions of algorithms interpretable by investigating what would have happened under different circumstances. These varying circumstances can be manipulations on the algorithm to assess its causality. In this thesis, I include the causal interpretability mechanism called the Average Causal Effect (ACE) into the training of a neural net. To assess the causality of the model, the ACE uses so-called interventions to manipulate the neural net. The goal is to determine more causal weights and biases during the model training. Using this approach, I evaluate whether including a causal interpretability mechanism as a regularization increases the overall causality of the model. Moreover, I investigate whether this improvement in causality also impacts the generalization ability of the neural net. The developed approach is compared to a standard neural net as well as L1 and L2 regularized neural nets. Furthermore, I conduct these experiments with well-balanced datasets, datasets with a prior probability shift, and datasets with a covariate shift. For all datasets, the results show that the presented causal regularization approach is able to improve the overall causality of the neural net. However, the distribution of the shifted training data highly affects the generalization ability. With an increasing variance of the distribution, the developed approach shows significantly lower test Mean Squared Errors than for training data with less variance. This is because the interventions applied by the ACE depend on the variance of the training data distribution.

      Email:-
      Status:

      FINISHED

      Supervisor:Matthias Kissel
    • Approximative Sparse Factorization of Neural Network Weight Matrices
      Student:Michael Brandner

      Abstract: In modern image processing applications, Convolutional Neural Networks (CNNs) are indispensable. Especially in the domains of object classification and face recognition, CNNs achieve impressive results. However, the increasingly accurate predictions are accompanied by ever-larger networks and consequently more computations. The number of operations required to compute the matrix vector product of a dense matrix M ∈ N N ×N in the fully connected layer scales with O (N 2 ) , mainly responsible that networks like VGG19 require 19.6 billion Floating-point Operations (FLOPs) to evaluate a single image. This thesis investigates the factorization of weight matrices, of the fully connected layer, into a product of sparse matrices, which potentially reduces the order of operations needed to the subquadratic domain. Consequently, the number of operations required for inference and thus, resource consumption is reduced. I examine three approximation algorithms, namely Butterfly factorization, sparse EigenGame, and Flexible Approximate MUlti-layer Sparse Transform ( F AμST ). The approaches are compared regarding the sparseness of their approximation and the approximation error. Furthermore, weight matrices of pre-trained Convolutional Neural Networks are factorized and compared regarding their prediction accuracy after approximation. 
      The best performance in terms of approximation error of the matrix and subsequent prediction accuracy was achieved by F AμST . F AμST was able to make sufficiently accurate predictions with only 20 % of the parameters. Where sufficient accurate means that the prediction accuracy drops by only 1 %. For similar results, the other algorithms needed 3 % (Butterfly) and 18 % (sparse EigenGame) more computations than the original matrix-vector product. 
      The experiments show that Approximative Sparse Factorization (ASF) of the weight matrices can significantly decrease resources consumption without deteriorating the accuracy of the predictions too much. This can enable complex computer vision algorithms to be used on devices with low computational resources or time-critical systems.

      Email:-
      Status:

      FINISHED

      Supervisor:Matthias Kissel
    • Algorithms for Matrix Approximations with Time Varying Systems
      Student:Stephan Nüßlein

      Abstract: There are different approaches to approximate matrices using structured matrices to reduce the computational cost of matrix-vector multiplications. A possible structure are sequentially semiseparable matrices, that describe the input-output behavior of time varying systems. If time varying systems are used to approximate weight matrices from neural networks, structural parameters have to be determined. In this thesis two algorithms to obtain the structural parameters are described. The first refines an initial segmentation by optimizing the input and output dimensions of the system. The second algorithm recursively splits the subsystems in a way that makes it possible to recover permuted sequentially semiseparable matrices. In experiments, the algorithms were able to recover the structure of simple test matrices. When used to approximate weight matrices from neural networks, the algorithms were able to reduce the computational cost of the matrix approximation compared to a naive approximation.

      Email:-
      Status:

      FINISHED

      Supervisor:Matthias Kissel
    • GAN-Based Differentially Private Publication of Vertically Partitioned Data

      Student/in

      Vorname: Yufei

      Name: Zhang

    • Deep Learning for Sound-to-Light Automation in Stage Lighting Applications
      Student:Manuel Hils

      Abstract: 

      Stage lighting applications usually require an operator who programs the dependency between sound and light. The industry is working on an improved sound-to-light automation whereas the scientific research in the field is limited. This thesis analyzed how sound-to-light automation can benefit from deep learning.

      Therefore, three different data sets were created: A synthetic data set where a light is pulsing with the beat, a trace of a basic real world light show and an extraction of control signals from a video of a professional light show. 

      The proposed model consists of a convolutional recurrent neural network (CRNN) that receives a log-Mel spectrogram together with its positive difference as input features. The loss function takes the absolute values of hue and brightness into account, as well as their change compared to the previous sample. Four additional metrics support the evaluation of the performance.

      The results showed that learning a direct dependency between low level music features and the lighting output via a CRNN is a promising approach for sound-to-light automation. 
      The first two data sets turned out to be more valuable than the third for this task. While the model learned the general structure of the first data set, especially the second revealed limitations that require further research. The relation between sound and light was not sufficient in the third data set.

      Email:manuel.hils@tum.de
      Status:


      Supervisor:Stefan Röhrl
    • Transformer-based Inner Speech Recognition from EEG data
      Student

      Unknown User (ga27tum)

      Emailragshanan.ragunathan@tum.de
      Status

      STARTING

      SupervisorMarisa Ripoll
      Abstract
    • Controllable Generation of Physically Based Rendering Materials from Albedo Texture using Generative Models
      Student:Max Gittel

      Abstract: Over the past decades, significant advancements have been made in the field of
      3D computer graphics, with ongoing research continuing to push the boundaries
      in this area. Simultaneously, the demand for 3D graphics applications is increas-
      ing across various industries, including film, video games, virtual and augmented
      reality (AR/VR), simulations, and visualizations.
      A key aspect of realistic 3D rendering is Physically Based Rendering (PBR),
      which simulates material shading with physically accurate properties. Most PBR
      materials rely on multiple textures such as albedo, roughness, metallic, height, and
      normal maps.
      This work explores whether a material’s full texture set can be generated solely
      from the artist-created albedo texture, aiming to automate parts of the workflow
      while maintaining artistic control. Two approaches are tested: a generative adver-
      sarial network (GAN) trained from scratch and a method based on a pre-trained
      Stable Diffusion model. The results show that the GAN approach tends to produce
      more consistent and structurally reliable textures, closely adhering to the albedo
      map, whereas Stable Diffusion fills in missing details and introduces creative vari-
      ations but offers less precise control and rarely preserves the original structure.

      Email:ge69fuv@mytum.de
      Status:

      RUNNING

      Supervisor:David Fresacher 
    • Human-AI Interaction in Programmatic CAD Model Generation
      Student: Christopher Lyko
      Abstract:

      Programmatic CAD systems allow users to define precise and reusable models through code, but they also require users to combine geometric reasoning with knowledge of a scripting language. Large Language Models create new opportunities to make programmatic CAD more accessible by translating natural-language design intentions into executable CAD scripts. However, AI-supported CAD generation does not remove the user from the design workflow. Generated results may be incomplete, incorrect, or misaligned with the user's intent, and design itself remains an iterative process of specifying requirements, creating alternatives, evaluating outcomes, and refining decisions.

      Supporting this workflow requires attention to the forms of interaction through which users inspect, refine, and steer AI-generated results. Human-AI interaction patterns are therefore analyzed across AI-supported CAD modeling, AI-supported programming, and co-creative design systems to evaluate their relevance for programmatic CAD workflows. Based on this analysis, a selected subset of interaction patterns is implemented in a prototypical environment for AI-assisted OpenSCAD modeling. The prototype explores how users can generate, inspect, refine, understand, and steer CAD models in collaboration with an AI system while retaining control over code, parameters, and the rendered model.

      The prototype is evaluated in a user study with beginner and experienced users. The study examines how the implemented interaction patterns affect the co-creation process, including the perceived quality of the resulting models, usability, perceived control, trust, and user satisfaction. A structured perspective on Human-AI interaction patterns for programmatic CAD is developed, leading to design implications for future AI-assisted CAD systems.
      Email: christopher.lyko@tum.de
      Status:

      STARTING
      Supervisor: David Fresacher 

  • No labels