Kolloquium am 25.11.2020, 10:00 Uhr

Deep learning is considered to be a breakthrough technology for many tasks in and outside
of computer vision. In order to achieve good prediction results, models need to be trained
on a sufficient amount of labeled data. This thesis presents Spherical-Sub-Sampling (SSS),
which combines sub-sampling and data augmentation for 3D point clouds in one step. SSS
subdivides a dataset and reuses some data points in multiple local environments. The
thesis provides an independent theoretical framework for SSS and explains the parameters
involved.
This is followed by multiple experiments, which evaluate the impact of SSS on semantic
segmentation of PointNet++ models that are trained on scans of crop fields. For this cause,
Rothamsted Research and Fraunhofer IIS provided 3D point cloud datasets of real wheat
field segments. First, the SSS parameters are tailored to this specific use case and then
multiple tests are conducted in realistic settings. The results suggest that SSS can boost
semantic segmentation performance particularly in cases, where available training datasets
are of limited size. It is also shown that running SSS multiple times cannot substitute the
benefits of more real data being available.