Research Article
Optimized PointNet for 3D Object Classification
@INPROCEEDINGS{10.1007/978-3-030-36402-1_29, author={Zhuangzhuang Li and Wenmei Li and Haiyan Liu and Yu Wang and Guan Gui}, title={Optimized PointNet for 3D Object Classification}, proceedings={Advanced Hybrid Information Processing. Third EAI International Conference, ADHIP 2019, Nanjing, China, September 21--22, 2019, Proceedings, Part I}, proceedings_a={ADHIP}, year={2019}, month={11}, keywords={Point cloud PointNet Object classification Center loss}, doi={10.1007/978-3-030-36402-1_29} }- Zhuangzhuang Li
Wenmei Li
Haiyan Liu
Yu Wang
Guan Gui
Year: 2019
Optimized PointNet for 3D Object Classification
ADHIP
Springer
DOI: 10.1007/978-3-030-36402-1_29
Abstract
Three-dimensional (3D) laser scanning technology is widely used to get the 3D geometric information of the surrounding environment, which leads to a huge increase interest of point cloud. The PointNet based on neural network can directly process point clouds, and it provides a unified frame to handle the task of object classification, part segmentation and semantic segmentation. It is indicated that the PointNet is efficient for target segmentation. However, the number of neural network layers and loss function are not good enough for target classification. In this paper, we optimize the original neural network by deepen the layers of neural network. Simulation result shows that the overall accuracy increases from 89.20% to 89.35%. Meanwhile, the combination of softmax loss with center loss function is adopt to enhance the robustness of classification, and the overall accuracy is up to 89.95%.
