One-Class Classification of 3D Point Clouds Using Dynamic Graph CNN

Anomaly detection is defined as a binary classification of normal and abnormal samples in a given data. In this task, anomalous samples are considered very rare and thus they are not used in training the deep learning model. Anomaly detection is thoroughly studied in 2D images, but 3D applications are not common yet. In this work, we combine studies in the field of 3D point cloud and anomaly detection and present a one-class classification method for 3D data. This method uses a pretrained Dynamic Graph Convolutional Neural Network (DG-CNN) to extract features from 3D point cloud data, and then fits a separate multivariate Gaussian distribution for each class. In the test phase, features are extracted from the test samples, and the distances to the distribution obtained in training indicate how anomalous the sample is for that class. Our method achieves 94.2% AUROC accuracy for the given setup. This is the first attempt to handle 3D point cloud classification as an anomaly detection case, and thus our findings and methods serve as a baseline for further studies.