Point Cloud Annotation Methods for 3D Deep Learning

Abstract

The domain of 3D Deep learning is growing rapidly as 3D sensor cost plunges and the perception capabilities these sensors can provide is continuously being extended. Dataset creation and annotation is a huge bottleneck in this field of work however, particularly in 3D segmentation tasks where every point in 3D space must be labelled accurately. This paper will review some creative ways of improving the data annotation process in terms of efficiency, accuracy and automatability. The review is comprised of two halves, firstly, annotation tools which have improved the user interface for pointcloud annotation are presented including works which use technologies such as virtual reality. Secondly, automation schemes which delegate as much of the work as possible to a machine while still giving the user insight and control over the process will be reviewed.