Eldorado - Repositorium der TU Dortmund

Ressourcen aus und für Forschung, Lehre und Studium

Bei diesem Service handelt es sich um das Institutionelle Repositorium der Technischen Universität Dortmund. Hier werden Ressourcen aus und für Lehre, Studium und Forschung gespeichert, erschlossen und der Öffentlichkeit zugänglich gemacht.

Hauptbereiche in Eldorado

Wählen Sie einen Bereich, um dessen Inhalt anzusehen.

Gerade angezeigt 1 - 5 von 9

Fach- und Universitätsübergreifendes 1320
Fakultäten 14068
Graduate School und Graduiertenkollegs 12
Hochschulverwaltung 1259
Sonderforschungsbereiche 1595

Aktuellste Veröffentlichungen

Leveraging data-based inductive biases in deep learning
(2025) Wagner, Stefan Sylvius; Harmeling, Stefan; Risse, Benjamin
Deep learning has revolutionised pattern recognition and decision-making, yet its success crucially depends on the inductive biases embedded in models implicit assumptions that guide learning toward useful solutions. These biases determine not only what a model can learn, but also how efficiently it can do so. Despite their importance, the sources and effects of inductive biases are rarely viewed in a more general manner. This dissertation investigates how inductive biases can be systematically identified and leveraged to improve learning across diverse domains. It formalises a distinction between two complementary kinds of inductive biases: (i) model-based inductive biases, which are encoded in architectures, loss functions, and regularization terms; and (ii) data-based inductive biases, which derive directly from structural or statistical regularities in the data. While the former has driven much of modern progress from convolutional networks to attention-based models, the latter remains comparatively underexplored despite its potential to capture task-specific invariances that are difficult to encode architecturally. The central contribution of this thesis is to define data-based inductive biases, which can be used to complement and enhance model-based inductive biases to create more robust and generalisable models. For this, a framework is introduced to describe how data-based inductive biases can be extracted and integrated into existing models, enhancing learning without the need for architectural redesign or extensive supervision. Latent variable modeling plays a key role in this framework, providing a principled means of discovering and encoding data structure as an additional learning signal. The chapters in this thesis demonstrate this principle across multiple domains. Cyclophobic Reinforcement Learning introduces a hand-crafted data-based inductive bias via hierarchical latent spaces over environment views, improving exploration by discouraging cyclic trajectories across these hierarchical perspectives. The Power of LLM-Generated Synthetic Data for Online Political Discussions uses synthetic data generated by large language models to define a controllable latent space, identifying ambiguous real-world samples that can be removed to reduce annotation cost or manually labelled to refine the model’s decision boundary. Just Cluster It: An Approach for Exploration in High Dimensions Using Clustering and Pre-Trained Representations shows how clustering DINO representations with Gaussian mixture models enables learning a latent density model over the state space, yielding more informative pseudo-count estimates for intrinsic motivation in 3D environments. Oh-A-DINO: Understanding and Enhancing Attribute-Level Information in Self-Supervised Object-Centric Representations demonstrates how augmenting DINOv2 representations with object-centric latent features learned through a Variational Autoencoder mitigates limitations of foundation models in representing non-geometric features such as colour and texture. Across these applications, the integration of data-derived structure consistently enhances performance and interpretability, showing that data-based inductive biases are a fundamental component to model learning.
Avatar scanning and personalization from consumer devices
(2026) Menzel, Timo; Botsch, Mario; Thies, Justus
This thesis presents methods for creating and personalizing realistic mesh-based avatars from smartphone input data. In the first part, we present an avatar reconstruction method that enables non-expert users to generate high-quality avatars using only consumer hardware, i.e., smartphones. To the best of our knowledge, this is the first fully automated end-to-end system that integrates guided smartphone capture with automated reconstruction to generate personalized, realistic, and VR-ready full-body avatars without requiring additional user input or preprocessing. Our reconstruction system consists of an iOS application that visually guides users through the scanning process and a server that automatically computes realistic avatars using photogrammetry, landmark detection, and template fitting. While this setup allows accessible avatar creation, the system is prone to motion-induced errors during scanning. We address this by reducing the effects of micro-motion at the photogrammetry stage and adapting the reconstruction method accordingly. Specifically, we compute partial point clouds from sub-scans using photogrammetry, incorporate silhouette constraints to improve fitting robustness, and apply inverse rendering to refine the texture on a per-image basis. The resulting avatars accurately match the scanned person's shape. However, animating the person's face is only possible with generic blendshapes. While this is sufficient to create new facial expressions, the resulting facial expression may lead to incongruent social cues. To address this, the second part of this thesis focuses on the personalization and animation of facial expressions. Building on the example-based facial rigging (EBFR) approach by Li et al. (2010), we introduce a method for personalizing blendshapes of realistic 3D avatars using smartphone input data. We developed an application to automatically capture example facial expressions. From these, we compute $52$ ARKit-compatible blendshapes using a position-based EBFR formulation and seamless partial deformation transfer. The resulting personalized blendshapes enable users to create faithful face animations. Finally, we evaluate the plausibility and naturalness of these facial animations by comparing different blendshape animation methods, including synthesized, tracked, and static (non-animated) facial expressions.
Benchmarking of algorithms on extended finite state automata
(2026) Dierl, Simon; Howar, Falk; Lamprecht, Anna-Lena
Extended finite state automata are a powerful formalism that can model many complex systems in a finite representation and remain amenable to analysis. A wide variety of algorithms interacts with these automata: for example, automata learning algorithms derive them from black-box systems, while model checking algorithms can verify their properties. When developing such algorithms, it is essential to perform benchmarking, i.e., analyze the algorithm’s performance on models of such machines. The choice of benchmark directly impacts the quality of the evaluation, with better benchmarks allowing for well-founded conclusions that are not restricted to, e.g., models of specific types of system. However, sources for benchmarks for extended finite state automata are limited. Only a single benchmark library and only simple synthesis algorithms are described in the literature. It is not clear what, given these restrictive circumstances, a good selection of benchmarks for a study is and to what extent choice of benchmarks ultimately impacts the strength of the study’s results. This thesis addresses this problem by analyzing limits in the availability and use of benchmarks in both extended finite state automata and finite state automata research and identifying the causes of these limits. We then address these causes and improve upon the current state of the art in benchmarking by defining two synthesis algorithms for extended finite state automata. We also tackle the lack of benchmark libraries by introducing recombinable benchmarks and tools, which simplify the creation and reuse of such libraries. By analyzing prior works and reflecting on the benchmarking performed in four case studies, we argue that the optimal approach to benchmarking is the combination of synthetic automata with well-curated, constantly-updated benchmark libraries. We successfully apply two novel synthesis algorithms in case studies and show that their benchmarks enable well-founded, strong conclusions. Finally, we retroactively analyze our approach to recombinability on our case studies and show its practicality in terms of effort required.
Amtliche Mitteilungen der Technischen Universität Dortmund Nr. 25/2026
(Technische Universität Dortmund, 2026-10-13)
Monolithic convex limiting and implicit pseudo-time stepping for calculating steady-state solutions of the Euler equations
(2024-12-18) Moujaes, Paul; Kuzmin, Dmitri
In this work, we use the monolithic convex limiting (MCL) methodology to enforce relevant inequality constraints in implicit finite element discretizations of the compressible Euler equations. In this context, preservation of invariant domains follows from positivity preservation for intermediate states of the density and internal energy. To avoid spurious oscillations, we additionally impose local maximum principles on intermediate states of the density, velocity components, and specific total energy. For the backward Euler time stepping, we show the invariant domain preserving (IDP) property of the fully discrete MCL scheme by constructing a fixed-point iteration that meets the requirements of a Krasnoselskii-type theorem. Our iterative solver for the nonlinear discrete problem employs a more efficient fixed-point iteration. The matrix of the associated linear system is a robust low-order Jacobian approximation that exploits the homogeneity property of the flux function. The limited antidiffusive terms are treated explicitly. We use positivity preservation as a stopping criterion for nonlinear iterations. The first iteration yields the solution of a linearized semi-implicit problem. This solution possesses the discrete conservation property but is generally not IDP. Further iterations are performed if any non-IDP states are detected. The existence of an IDP limit is guaranteed by our analysis. To facilitate convergence to steady-state solutions, we perform adaptive explicit underrelaxation at the end of each time step. The calculation of appropriate relaxation factors is based on an approximate minimization of nodal entropy residuals. The performance of proposed algorithms and alternative solution strategies is illustrated by the convergence history for standard two-dimensional test problems.