Hybrid Transformer-CNN Architecture for Efficient Point Cloud Classification with Attention Mechanisms

Abstract

This  paper  proposes  a  Transformer-based  point cloud  classification  method.  By  designing  two  key  modules:  a hybrid feature extraction module and a multi-head self-attention module,  efficient  feature  extraction  and  classification  of  point cloud data are achieved. The hybrid feature extraction module combines the advantages of convolutional neural network (CNN) and Transformer architecture, extracts local geometric features through  convolution  operations.  The  multi-head  self-attention module  captures  different  feature  relationships  in  the  point cloud  through multiple  self-attention heads  to further  enhance feature representation. Experimental results also demonstrate the effectiveness of our method. Our method achieves 91.6% mAcc and  93.3%  OA  on  ModelNet40  dataset;  on  the  ScanObjectNN dataset, it achieves 79.8% mAcc.