Qendrim Schreiber, a former student of the Mathematics program and staff member at HFT Stuttgart, has successfully completed his doctorate at Johannes Gutenberg-University Mainz with his doctoral colloquium.
DeThe title of his dissertation is: “Deep Learning Methods for Representation Learning with a Focus on 3D Geometric Data.”
The doctoral research was conducted within the framework of the GeoCADUp and iCity2 projects at HFT Stuttgart, in collaboration with Johannes Gutenberg-University Mainz. The work investigates representation learning in deep learning with the aim of improving classification and segmentation tasks, particularly for 3D geometric data. To this end, three central contributions are presented: a new loss function (Prototype Softmax Cross Entropy, PSCE), an architecture for processing point clouds (LocAL-Net), and an approach for the structured processing of large meshes (METNet).
The first part examines the classification of 3D objects, for example determining whether an object is a table or a chair. The classical loss function Softmax Cross Entropy is analyzed and interpreted in the context of contrastive learning and extended by PSCE, which enables a significantly better separation of classes in the feature space through freely selectable prototypes. The second part introduces LocAL-Net, which specifically captures local geometric structures in point clouds and effectively combines them with global contextual information. The third part presents METNet, which enables the segmentation of large city models represented as textured triangular meshes. This allows each triangle to be assigned, for example, to a house, a tree, or a vehicle. In METNet, geodesic neighborhoods in the triangular mesh are represented as structured tensors, enabling more efficient processing with neural networks and leading to significantly improved segmentation results.
We warmly congratulate Dr. Schreiber on successfully completing his doctorate!
