The core of the project is the application, extension and development of modern techniques from the field of machine learning for object segmentation from unstructured 3D point clouds. An important question is to decide how fine granular with classical approaches of semantic segmentation the individual elements can be determined as fully automated as possible. Thereby, the annotation effort for the training should be kept as low as possible. Furthermore, approaches are to be investigated and developed which make it possible to integrate expert or domain knowledge into the semantic segmentation. So-called "generative adversarial networks" for unsupervised modeling are one object of investigation. The research and development work also includes the elaboration of suitable forms of representation to capture the individual objects (e.g., skeletons, graphs). Finally, the developed methods will be applied to the application scenario "tree mapping from 3D LiDAR data" and it will be investigated to what extent the methods can be transferred to "building mapping from 3D stereo data".
Project runtime: 2017 - 2021
Spokeperson: Dr. Friederike Klan