Handling Data Scarcity Through Data Augmentation in Training of Deep Neural Networks for 3D Data Processing

Abstract

Due to the availability of cheap 3D sensors such as Kinect and LiDAR, the use of 3D data in various domains such as manufacturing, healthcare, and retail to achieve operational safety, improved outcomes, and enhanced customer experience has gained momentum in recent years. In many of these domains, object recognition is being performed using 3D data against the difficulties posed by illumination, pose variation, scaling, etc. present in 2D data. In this work, the authors propose three data augmentation techniques for 3D data in point cloud representation that use sub-sampling. They then verify that the 3D samples created through data augmentation carry the same information by comparing the iterative closest point registration error within the sub-samples, between the subsamples and their parent sample, between the sub-samples with different parents and the same subject, and finally, between the sub-samples of different subjects. They also verify that the augmented sub-samples have the same characteristics and features as those of the original 3D point cloud by applying the central limit theorem.