3D Point Cloud Classification
In recent years, with the gradual advancement of artificial intelligence algorithms and big data, various industries and a growing workforce have begun investing in artificial intelligence-related fields. Notably, 3D acquisition technology and hardware equipment have made significant strides, with the collected 3D data finding widespread applications in computer vision and artificial intelligence domains.
We focuse on two key characteristics of 3D point cloud data: disorder and invariance. The objective is to develop a model that achieves high accuracy with low complexity. To accomplish this, we employ a projection method, projecting 3D point cloud data into 2D point cloud data and converting it into a multi-angle RGB image for use as training data.
The proposed classification method, MPML-GCN, utilizes the ModelNet40 dataset under different views. It employs multi-view two-dimensional images as training data, incorporating a multi-level graph convolutional neural network and MVCNN network architecture. This approach is designed to hierarchically extract point cloud shape features and aggregate these features for the classification of 3D objects.
Through classification experiments conducted on the ModelNet40 dataset, this research demonstrates that the MPML-GCN classification method yields significantly higher accuracy compared to some methods that directly utilize 3D point cloud data.