Masterarbeit - Explainability for 3D Scene Understanding

The Fraunhofer-Gesellschaft (www.fraunhofer.de) is one of the world's leading organizations for application-oriented research. 74 institutes develop pioneering technologies for our economy and society – more precisely, 30,000 people from engineering, science, administration, and IT. They know: Anyone who joins Fraunhofer wants to and can make a difference. For themselves, for us, and for the markets of today and tomorrow. At the Fraunhofer Institute for Physical Measurement Techniques IPM, we work every day to make optical technologies usable for specific measurement tasks. At the interface between science, research, and industry, we develop tailor-made solutions and stand for application-oriented research at the highest level. In the Object and Shape Detection department, mobile measurement systems are developed for surveying infrastructure ranging from roads and railways to complex indoor environments. The development process covers every step. From design and engineering to system integration and the automatic interpretation of measurement data. A cornerstone of these systems is the automated, AI-driven interpretation of 3D point cloud data, which demands high-quality, training datasets and idealy model interpretability. The primary objective of this thesis is to integrate explainability-driven approaches into the training pipeline. These could be based on 3D uncertainty estimation to pinpoint specific spatial regions or concept bottleneck models to inspect intermediate representations of the model. These are only examples, and the thesis is not strictly limited to these approaches. By applying the chosen methods to different domain-specific benchmarks, the resulting thesis will bridge the gap between model interpretability and dataset quality, delivering robust, efficient, and transparent 3D point cloud solutions.