Malte Stelzer, M.Sc.

Offene Abschlussarbeiten

A Modular Multimodal Framework for Evaluating Coherence and Consistency in Generative AI Outputs for RGB, LiDAR, and RaDAR Data

Generative AI models are increasingly capable of producing synthetic multimodal sensor data, including RGB images, LiDAR point clouds, and Radar signals. Such models are valuable for simulation, data augmentation, and training perception systems in autonomous driving. However, assessing the quality of generated multimodal data remains challenging, particularly with respect to cross-modal coherence and temporal or spatial consistency.

In this work, you will design and implement a modular neural framework to evaluate the coherence and consistency of generative AI outputs across RGB, LiDAR, and Radar modalities. The system will analyze both intra-modal quality and cross-modal alignment between different sensor representations. You will develop modular scoring components that can be independently applied to each modality while also integrating a fusion module to measure multimodal consistency.

The resulting framework should support multiple generative model architectures and operate as a general evaluation tool for multimodal synthetic data. The goal is to provide quantitative metrics and visualization tools that help researchers analyze and compare generative models used in autonomous driving and related applications.