Point Cloud Segmentation with Deep Reinforcement Learning

Abstract

The segmentation of point clouds is conducted with the help of deep reinforcement learning (DRL) in this contribution. We want to create interactive virtual reality (VR) environments from point cloud scans as fast as possible. These VR environments are used for secure and immersive trainings of serious real life applications such as the extinguishing of a fire. It is necessary to segment the point cloud scans to create interactions in the VR. Existing geometric and semantic point cloud segmentation approaches are not powerful enough to automatically segment point cloud scenes that consist of diverse unknown objects. Hence, we tackle this problem by considering point cloud segmentation as markov decision process and applying DRL. More specifically, a deep neural network (DNN) sees a point cloud as state, estimates the parameters of a region growing algorithm and earns a reward value. The point cloud scenes originate from virtual mesh scenes that were transformed to point clouds. Thus, a point to segment relationship exists that is used in the reward function. Moreover, the reward function is developed for our case where the true segments do not correspond to the assigned segments. This case results from, but is not limited to, the usage of the region growing algorithm. Several experiments with different point cloud DNN architectures such as PointNet [13] are conducted. We show promising results for the future directions of the segmentation of point clouds with DRL.

Acknowledgements

This project is sponsored by: German Federal Ministry of Education and Research (BMBF) under the project number 16SV8182 and 13FH022IX6. Project names: HIVE-Lab and Interactive body-near production technology 4.0 (German: Interaktive körpernahe Produktionstechnik 4.0 (iKPT4.0)).