This research uses PVCNN as a method to classify objects directly in 3D point clouds sourced from lidar, resulting in each point being attributed with a class label (e.g. powerline, roof). The ISPRS Vaihingen 3D semantic dataset was used for training and testing purposes. In addition to xyz point coordinates, geometric properties were calculated from the point cloud and added as attributes, used in training and inference. Four trials were undertaken to compare performances, based on F1-scores, using different combinations of engineered features (eg xyz+linearity; xyz+planarity). Surprisingly linearity was less useful than planarity in assisting the classification process even with linear powerline features. The method shows promise for its processing times on low powered GPUs. Future work will explore the impact on classification performance with additional engineered features, height above ground values per point, and adapt the spatial clustering algorithms used.