Comparative Study of Point Cloud Classification Using Deep Learning Neural Networks

Point cloud processing using neural network is challenging due to its irregular and unstructured nature. In recent development, two general types of deep learning neural networks can generate point cloud in canonical representation for good classification performance, i.e. point-based method and neighborhood-based method. Point-based methods process every single point with neural linking to find its representation, while neighborhood-based methods compute the local correlation between each point and its neighbors. To study the difference of both approaches, a comparative study on the benchmarking PointNet and Dynamic Graph Convolutional Neural Network (DGCNN) with several experiments are provided in this paper to highlight their advantages and limitation. The comparison between DGCNN and PointNet is made in terms of training time in an epoch, inference time, accuracy and the number of parameters on point cloud classification tasks using the ModelNet40 dataset. The confusion matrices are visualized to provide an in-depth comparison of inter-class and intra-class of objects classification. As a result, DGCNN achieves better overall accuracy and a lower number of parameters (92.87% and 3.47 million parameters) when compared to PointNet (90.05% and 1.81 million parameters). However, PointNet achieves a faster training time (32 seconds) and inference time (3 seconds) as compared to DGCNN training time (162 seconds) and inference time (11 seconds).