Recent Submissions
Domain adaptation methods for emotion and pain recognition via video games
(2024) Nasimzada, Jonas
Seeing the patient’s emotional and physical condition is crucial when designing patient-computer interaction systems. However, gathering large datasets in sensitive situations like filming a person in pain can be challenging and ethically questionable.
The primary aim of this study is to assess the possibility of using synthetic data as an alternative data source to create models capable of effectively recognizing patient pain. Initially, a synthetic dataset was generated as the foundation for model development. To maintain the relevance of the synthetically generated dataset’s diversity, a 3D model of real people was created by extracting facial landmarks from a source dataset and generating 3D meshes using EMOCA (Emotion Driven Monocular Face Capture and Animation) [1] [2]. Meanwhile, facial textures were sourced from publicly available datasets like CelebHQ [3] and FFHQ-UV [4].
An efficient pipeline was created for human mesh and texture generation, resulting in a dataset of 8,600 synthetic human heads generated in approximately 2 hours per perspective and texture. The datasets encompass varying facial textures and perspectives and total over 300 GB. This approach enhances gender and ethnic diversity while introducing perspectives from previously unseen viewpoints.
Combining the 3D models with the extracted textures created new characters with varying facial textures but identical facial expressions. The study aims to bridge the gap between synthetic data and real-world medical contexts using domain adaptation methods, like Domain Mapping. This approach eliminates the need for human participants and addresses ethical issues associated with traditional data collection methods.
Different combinations of datasets, encompassing various textures and perspectives, were utilized to train models and assess the feasibility of synthetic data for domain adaptation (Domain Mapping) with real human data as input video.
However, incorporating synthetic and real data leads to improved pain recognition capabilities. This combined approach can leverage the strengths of both real and synthetic datasets, resulting in a more robust and effective model for pain recognition.
Medienbrüche auflösen und Prozesse in gewachsenen Systemen durch die Nutzung von Schnittstellen harmonisieren
(2024) Dukino, Claudia; Pawlowicz, Daniel; Flex, Michael; Oberdörfer, Peter
Patterns of neural activity in response to threatening faces are predictive of autistic traits : modulatory effects of oxytocin receptor genotype
(2024) Zheng, Xiaoxiao; Zhou, Feng; Fu, Meina; Xu, Lei; Wang, Jiayuan; Li, Jialin; Li, Keshuang; Sindermann, Cornelia; Montag, Christian; Becker, Benjamin; Zhan, Yang; Kendrick, Keith M.
Autistic individuals generally demonstrate impaired emotion recognition but it is unclear whether effects are emotion-specific or influenced by oxytocin receptor (OXTR) genotype. Here we implemented a dimensional approach using an implicit emotion recognition task together with functional MRI in a large cohort of neurotypical adult participants ( N = 255, male = 131, aged 17-29 years) to establish associations between autistic traits and neural and behavioral responses to specific face emotions, together with modulatory effects of OXTR genotype. A searchlight-based multivariate pattern analysis (MVPA) revealed an extensive network of frontal, basal ganglia, cingulate and limbic regions exhibiting significant predictability for autistic traits from patterns of responses to angry relative to neutral expression faces. Functional connectivity analyses revealed a genotype interaction (OXTR SNPs rs2254298, rs2268491) for coupling between the orbitofrontal cortex and mid-cingulate during angry expression processing, with a negative association between coupling and autistic traits in the risk-allele group and a positive one in the non-risk allele group. Overall, results indicate extensive emotion-specific associations primarily between patterns of neural responses to angry faces and autistic traits in regions processing motivation, reward and salience but not in early visual processing. Functional connections between these identified regions were not only associated with autistic traits but also influenced by OXTR genotype. Thus, altered patterns of neural responses to threatening faces may be a potential biomarker for autistic symptoms although modulatory influences of OXTR genotype need to be taken into account.
“It’s just politics” : an exploration of people’s frames of the politics of mobility in Germany and their consequences
(2024) Sonnberger, Marco; Leger, Matthias; Radtke, Jörg
Background. The decarbonization of the mobility sector is one of the main challenges in the context of climate mitigation. In Germany, as in many other countries, policy measures aiming to make the mobility system greener frequently fail to produce substantial results, not least due to a lack of support by large sections of the general public. Policy measures directed at reducing car traffic in particular are often met with indifference and resistance. The question thus arises: what basis do citizens use to form their (often negative) opinions about sustainable mobility policies? As a conceptual starting point for our empirical analysis, we draw on the frame concept and focus on people’s frames of the politics of mobility. With “politics of mobility” we refer to everything people could consider as political with regard to mobility. We understand frames as culturally mediated patterns of interpretation that ultimately motivate and guide actions.
Results. Based on interviews and focus group data gathered in the region of the city of Stuttgart (Germany), we identify two dominant frames as well as combinations of these frames by which people make sense of the activities of political actors in the field of mobility. In one frame, which we labeled “politics-as-actor”, mobility politics are interpreted with reference to politics as some kind of monolithic abstract actor. In the other, which we labeled as “politics-as-staged-process”, mobility politics are portrayed as an interest-driven, opaque process that only purport to being democratic.
Conclusions. In terms of policy recommendations, we use our findings to derive suggestions for how to increase support for green mobility policies: transparent implementation of policy measures, pragmatic policy styles and the involvement of intermediaries.
Multi-human behavior prediction using vision language models
(2025) Panchal, Utsav
The ability to accurately predict multiple human motions and behaviors is crucial for mobile robots operating in human-populated environments. It is essential to incorporate the context of the scene and the states of objects within the environment because human behaviors are inherently influenced by their surroundings. Although prior research focuses primarily on predicting actions in single-human scenarios from an egocentric view, robotic applications require understanding multiple human behaviors from a third-person perspective. In contrast, there are fewer pre-existing datasets that captures multi-human behavior, especially from a third person perspective. The gap in data availability further complicates the development of accurate and efficient prediction methods for real world applications. This thesis addresses the problem of forecasting actions of multiple humans within a scene from a third person’s point of view. By leveraging Vision Language Models (VLMs) and Scene graphs, this thesis proposes a framework that is capable to predict multiple-human behavior in an indoor environment. Due to a lack of suitable dataset for multiple human behavior prediction, this thesis also fine-tunes open source VLMs with synthetic human behavior data and evaluates the resulting models on both synthetic sequences and real-world video recordings to assess their generalization capabilities. Additionally, this thesis also outlines the process of generating synthetic data generated by using a photo-realistic simulator. This thesis presents VISTA, which stands for Vision And Scene Aware Temporal Action Anticipation, a fine tuned VLM which is capable to predict human behavior up-to 5 seconds in a single-shot manner. This work also details a fine-tuning pipeline for VISTA utilizing methods such as Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO) resulting in an 13% improvement over existing methods. In the end, this thesis also presents several ablation studies to examine different components of the framework and to understand the factors influencing the behavior prediction.
Optimising the generation performance for multimodal diffusion models using reinforcement learning
(2024) Gaude, Justus
This bachelor thesis explores the field of multimodal data synthesis, focusing specifically on the generation of high quality image-text pairs within the UniDiffuser framework. While the UniDiffuser framework has proven its efficiency in generating joint samples along a linear path, where the timesteps of the modalities are uniformly discretized. This study questions whether alternative paths could potentially offer better outcomes in terms of both quality and efficiency. To address this inquiry, hypotheses are formulated, an environment is developed, action space, and state spaces are defined. Through the training of reinforcement learning agents and the use of evaluation metrics, this research attempts to find alternative paths that are more computationally efficient and produce higher quality image-text pairs. Ultimately, this study aims to advancing the state of the art in multimodal data generation.
Propädeutik der Technikwissenschaften : Doktorandenseminar: Wissenschaftstheorie der Technikwissenschaften
(Stuttgart : Universität Stuttgart, Institut für Industrielle Fertigung und Fabrikbetrieb, 2025) Erlach, Klaus
Technikwissenschaften im Wissenschaftssystem
Vom Forschungsprozess zur Gliederung: Forschungsreise
Technikwissenschaftliche Erkenntnis: Begriff & Hypothese
Modelltheorie und Systemtheorie
Forschungsfrage und Themenfindung
Evaluation der Zielerreichung: Verifikation & Validierung
Wissenschaftstheoretische Positionierung
Towards advancing modular AI planning systems : insights from the PlanX toolbox redesign
(2025) Six, Daniel
As ai planning systems grow in complexity and are deployed in increasingly diverse application contexts, existing frameworks often struggle with modularity, reusability, and tool interoperability. These limitations hinder the integration of heterogeneous planning tools and reduce the adaptability of planning systems to evolving requirements. This thesis addresses these issues by redesigning the PlanX Toolbox, an open-source framework for composing ai planning functionalities based on soa and cbse principles.
Through a detailed architectural analysis of the original PlanX system, several constraints were identified, including rigid service interactions, limited extensibility, and tightly coupled parsing, converting, and planning components. In response, a redesigned architecture was developed featuring a modular Generation Unit that unifies conversion and plan generation while maintaining parsing as an independent component. This restructuring improves the separation of concerns and supports more flexible composition of planning workflows. The redesign also introduces standardized message structures and dynamic planner selection, simplifying the integration of new components.
The evaluation includes the integration of the Fast Downward planner, which revealed both the potential and the limitations of reusing standardized input formats like pddl across diverse planning backends. While conversion and generation could be modularized effectively, parser reusability remained constrained by tool-specific assumptions and preprocessing steps. These findings highlight key insights into the practical challenges of building modular ai planning systems: true composability requires not only standardized interfaces but also architectural awareness of planner internals.
By proposing and validating a refined system architecture, this thesis provides actionable design patterns for enhancing maintainability, extensibility, and integration in ai planning systems. The findings offer both theoretical and practical contributions to advancing modular ai planning frameworks.
Entwicklung eines Datenprodukt Lifecycles zur Verbesserung des Datenproduktmanagements
(2025) Herrmann, Lisa
Daten sind einer der wertvollsten Rohstoffe der heutigen Zeit, werden jedoch von vielen Unternehmen nicht optimal genutzt, obwohl sie in großen Mengen vorliegen. Ein erleichterter Zugriff und Umgang mit Daten sind essenziell, um deren Potenzial auszuschöpfen. Besonders ein domänenübergreifender Austausch fördert die Zusammenarbeit und eine effektivere Nutzung. Data Mesh stellt als dezentrales Konzept einen vielversprechenden Ansatz für das Datenmanagement dar, indem Daten in Form von Datenprodukten von den jeweiligen Domänen bereitgestellt werden. Ein strukturierter Lifecycle kann das Management dieser Datenprodukte optimieren, indem klare Zuständigkeiten der beteiligten Rollen definiert werden. Bestehende Konzepte wie Produkt-Lifecycles oder Daten-Lifecycles werden im Zuge einer Anforderungsanalyse als unzureichend für die Anwendung im Bereich Data Mesh befunden, da diese Konzepte wesentliche Anforderungen der Rollen nicht vollständig abdecken. Daher wird in dieser Arbeit ein neuer Datenprodukt-Lifecycle entwickelt, der sich in die fünf Hauptphasen Discovery, Design, Development, Deployment und Delivery gliedert. Ergänzend umfasst der Lifecycle andauernde Tätigkeiten wie Maintenance, Monitoring und die Einhaltung von unternehmensweiten Vorgaben. Zudem werden Implementierungsstrategien für die Umsetzung aufgestellt, und zur Veranschaulichung wird das Konzept prototypisch als grafische Benutzeroberfläche implementiert. Die Evaluation zeigt, dass der Lifecycle die beteiligten Rollen im Umgang mit Datenprodukten unterstützt, indem eine klare Differenzierung der Verantwortlichkeiten ermöglicht wird und damit zur Verbesserung des Datenproduktmanagements beigetragen wird.
Deep fair clustering with multi-objective handling
(2025) Saha, Ankita
Clustering is an unsupervised learning technique widely used in various critical domains, such as heathcare and finance, yet fairness remains an underexplored challenge especially in the field of deep-learning based clustering. Traditional methods often reinforce biases present in the dataset, leading to ethical concerns in critical applications This thesis proposes Deep Fair Clustering with Multi-Objective Handling (DFC-MOH), a deep learning framework integrating both group level and individual Level fairness constraints with clustering quality. By incorporating balance loss for group fairness and individual fairness loss into a combined loss function, DFC-MOH enables flexible trade-offs between both fairness constraints and clustering performance. Extensive experiments on synthetic and real-world datasets demonstrate the effectiveness of our approach in achieving fair and high-quality clustering with reasonable scalability.