TCPNet: A 3D Point Cloud Classification Model Based on the Advanced CNN and Transformer

Transformer and Convolutional Neural Network (CNN) are currently two important models in the field of deep learning. Among them, Transformer has strong global perception ability but weak local perception ability, and CNN precisely compensates for this. CNN has strong local perception ability, so in order to combine the advantages of both, we propose a 3D classification model TCPNet based on Transformer and CNN. We first use Transformer to extract the global features of the point cloud, and then concatenate a CNN network to enhance the detail perception ability, so that the network can better understand the 3D point cloud. The experimental results on the ModelNet40 and ScanObjectNN datasets show that our method can effectively achieve 3D classification tasks. On ModelNet40, an accuracy of 93.6% can be achieved. On the most difficult variant PB-T50-RS of ScanObjectNN, an accuracy of 83.93% can be achieved.