Data augmentation by generating 3D adversarial point cloud

Abstract

The use of 3D data in many applications has gained traction in recent years due to the availability of inexpensive 3D sensors such as Kinect and LIDAR. Object iden tification has gained popularity, and it is now done with 3D data to overcome the challenges posed by position, expression, lighting fluctuations, and occlusions in 2D data. Though 3D object recognition is more accurate, the availability of 3D data is limited, causing overfitting in Deep Neural Networks and making processing more difficult, requiring more space and time. To circumvent this issue, the variability of 3D data must be raised by using data augmentation to expand the quantity of avail able data. Adversarial examples are carefully constructed instances that drive Deep Neural Network models to make incorrect predictions. In this report, we present a way of doing data augmentation by creating 3D adversarial examples. We use random and stratified techniques to generate augmented data for each 3D sample of a subject. The augmented data is trained using PointNet augmented with the siamese model, with the generated augmented data from a sample labeled as a single class. We use the trained model to implement the Adversarial Point Perturbation technique and generate the perturbed data. We then compare the Iterative Closest Point Registration Error between a pair of perturbed samples belonging to the same class and their respective parent augmented samples pair, and between a pair of per turbed samples belonging to different classes and their respective parent augmented samples pair to ensure that the perturbed data created carries the same information as the parent augmented data. Keywords: Data Augmentation, Adversarial Point Clouds, PointNet, Siamese Network, Adversarial Point Perturbation, ICP