Abstract
Point clouds are a common representation of 3D scenes, and labelled point clouds are necessary as input to many machine learning systems. Current labelling tools, however, are predominantly 2D. A 3D interface would be more natural fit to the task, and we investigate Virtual Reality as a mechanism for point cloud labelling. In contrast to previous studies we find that the choice of 2D or 3D interface is not the determining factor for labelling speed or accuracy. The nature of the task is important, with some tasks being better suited to the 2D or 3D tools.