Neural Processes for Optimal Sensor Placement
Machine learning models for optimal ocean measurement locations using uncertainty-aware Neural Processes.
University of Rostock · Institute for Visual and Analytic Computing
Marine Data Science
Research group working at the intersection of machine learning, marine data, sensor systems, probabilistic modeling, and visual computing.
Member of the CORE Network (Cognitive Robotics in Europe)
Research
Machine learning methods for marine data, sensor systems, medical imaging, tabular data, and human activity recognition.
Ultrasound Image Analysis
AI-driven real-time 3D ultrasound analysis for diagnostic and therapy guidance.
Tool Use in Large Language Models
Methods that let language models call external tools reliably, from function calling and MCP to agents that act on real software systems.
Deep Learning for Tabular and Categorical Data
Deep neural network models for tabular and categorical data beyond tree-based baselines.
Projects
Funded research projects.
ADAPT-AI
ADAPT-AI: Intelligent Adaptation of Machine Learning Models for Time-Series Data
Show2Instruct
Multimodal generative AI for turning natural-language interaction with complex software systems into machine-processable instructions.
Data4Sim
Sensor-based capture and simulation of manual logistics processes for testing operational changes before implementation.
For students
Theses
We are always looking for motivated students to work on Bachelor and Master thesis topics.
- medical imaging
Open Applications in Medical Image Computing
Students interested in pursuing a thesis related to medical applications are welcome to contact Daniel Wulff.
- Robotics
RoboBIM-Sim: Semantically Articulated USD Environments from IFC Building Models
Automated IFC-to-USD pipeline producing physics-ready, semantically annotated simulation environments for Embodied AI.
- Graph Learning
- BIM
GNNs for BIM: Learning Relationships in IFC Models
Learning the recurring relationships between objects in valid IFC models with graph neural networks.
- Computer Vision
Scan2BIM-Data: Automated IFC Generation for Text-to-BIM
3D semantic segmentation and primitive fitting to convert point clouds into parametric IFC datasets for Text-to-BIM.
- Information Retrieval
- BIM
BIM-RAG for LLM-Based Building Design
Building a searchable database of BIM components, standards, and code documentation, and testing whether retrieval improves LLM-based BIM editing.
- marine
Deep learning-based classification of phytoplankton in imaging flow cytometry data
- Embodied AI
Spatial LLM Agents for Zero-Shot Visual Positioning
Feeding point cloud semantics directly into multimodal LLMs for abstract spatial reasoning and zero-shot routing.
- Embodied AI
Tiered Spatial Memory: Detail on Demand for Robotic Manipulation
A hierarchical object graph that generates high-fidelity geometry only when a manipulation task requires it.
- Computer Vision
Code-as-Inverse-Graphics: Benchmarking Spatial Understanding in Vision-Language Models
Benchmarking VLM spatial reasoning by tasking agents to reconstruct Blender scenes as executable bpy code from images.
- machine learning
Methods for Pseudo-Label Selection for Student-Teacher-based Domain Adaptation
Methods for selecting pseudo-labels in semi-supervised learning scenarios to improve model performance with unlabeled data.
Teaching
Current lectures and student projects.
EgoProject 2026
Summer semester group project on spatial understanding in egocentric data.
Artificial Intelligence (UBB Cluj-Napoca)
Artificial Intelligence teaching material for UBB Cluj-Napoca.
RoOT
RoOT Summer School material in Rostock.
People
Current team members.
Feyza Eksen
PhD student
Corinna Burmeister
Team Assistant
Peter Eschholz
Systems Engineer