In this thesis, we present a pipeline of how to train an object detector for a point cloud if there are no ready-to-use data for the specific objects to train on except their own point cloud. Going beyond producing synthetic data with the object poses only, instead, we choose the augmentation approach and enrich the existing benchmark dataset with the existing object point clouds to provide a better data variability, thus, hopefully, im- proving the capabilities of the object detector in realistic highly-cluttered scenes as well as narrowing the domain gap. At the same time, point cloud augmentation is a time- saving alternative to producing more complex synthetic scenes since it does not involve computation-demanding physical simulations. The benchmark dataset chosen for the point cloud augmentation contains a sucient amount of indoor point clouds of the furnished rooms out which we chose around 500 scenes for the augmentation and subsequent training of our object detector. Using the available point clouds, we extract enough horizontal planes as the locus for the possible insertions of the object point clouds. To provide more variability in the data, we insert objects in random poses so that their lowest point is the contact point with a randomly chosen location. By avoiding insertions that would lead to occlusions with either other inserted objects or with the initial scene and by performing hidden point removal, we ensure the realistic appearance of the obtained augmented point cloud. Though quite straight-forward, this, nevertheless, e ective approach yielded, for some cases, considerably good results when trained on our object detector. In particular, it was able to distinguish the objects from the background well enough and likewise showed prom- ising results on the object location not only on the validation data but also when tested on the real data captured in some of the laboratories and oce environment of Fraunhofer IPA. The angular component of the object pose remains, however, still challenging.