Master Thesis – Enhanced Reasoning in Vision-Language Models Using Structured Visual Representations
Core
Investigate how structured visual information from surveillance metadata can be integrated into Vision-Language Models to improve spatial, object-level, and temporal reasoning.
Role type
Master thesis researcher (Machine Learning / Computer Vision)
Builds
Fine-tuned Vision-Language Models using structured visual representations
Domain
Computer Vision, Machine Learning, Surveillance Analytics
Deliverable
research
Required skills
Machine Learning, Vision-Language Models, Computer Vision, Programming, Instruction Fine-Tuning, Parameter-Efficient Fine-Tuning (PEFT)
Preferred skills
Experience with open-source VLMs, knowledge of surveillance system metadata
Technologies
LoRA, PEFT, Open-source VLMs, Surveillance metadata (bounding boxes, trajectories, velocity vectors)
Responsibilities
Investigate state-of-the-art methods for integrating structured visual information, explore use of existing metadata, fine-tune VLMs, evaluate model performance on spatial and temporal reasoning tasks, compare models with and without structured information, investigate interpretability.
Seniority
Master student level
