Lightweight Network for Point Cloud Classification Based on Gridding and Surface Features Encoder
Point clouds carry rich three-dimensional features,and their classification problem has always been a hot topic in the field of deep learning.The accuracy of existing point cloud classification networks is already relatively ideal,but the parameter and computational complexity are too large,which is not conducive to deployment in practical scenarios.A lightweight point cloud classification network,GridPoint,is pro-posed to address this issue.Firstly,design a point cloud gridding module,which divides the grid area based on the coordinate position of the points;Then expand the higher-order term function of the coordinates,encode the surface features of the original point cloud,and enhance the expression of contour features;Finally,two rounds of global pooling are used to extract local features and aggregate global features.Perform clas-sification and ablation experiments using the classic point cloud dataset ModelNet40,ShapeNetCore,and the real dataset ScanObject NN.The ex-perimental results show that the classification accuracy of GridPoint is close to mainstream networks such as PointNet++,with a difference be-tween 0.3%and 2.3%;The network parameters and computational complexity are 0.11 M and 0.05 G,respectively,which are reduced by more than 81.7%and 88.9%compared to mainstream networks.They have significant advantages in lightweight and have good practical value.
deep learningpoint cloud classificationlightweight networkpoint cloud griddingsurface feature encoder
国家科技期刊平台
NSTL
万方数据
