TY - GEN
T1 - 3D Point Cloud Classification Based on Convolutional Neural Network
AU - Lu, Jianrui
AU - Kang, Wenjing
AU - Ma, Ruofei
AU - Qin, Zhiliang
N1 - Publisher Copyright:
© 2022, ICST Institute for Computer Sciences, Social Informatics and Telecommunications Engineering.
PY - 2022
Y1 - 2022
N2 - With the development of science and technology, the requirements for 3D point cloud classification are increasing. Methods that can directly process point cloud has the advantages of small calculation amount and high real-time performance. Hence, we proposed a novel convolutional neural network(CNN) method to directly extract features from point cloud for 3D object classification. We firstly train a pre-training model with ModelNet40 dataset. Then, we freeze the first five layers of our CNN model and adjust the learning rate to fine tune our CNN model. Finally, we evaluate our methods by ModelNet40 and the classification accuracy of our model can achieve 87.8% which is better than other traditional approaches. We also design some experiments to research the effect of T-Net proposed by Charles R. Qi et al. on 3D object classification. In the end, we find that T-Net has little effect on classification task and it is not necessary to apply in our CNN.
AB - With the development of science and technology, the requirements for 3D point cloud classification are increasing. Methods that can directly process point cloud has the advantages of small calculation amount and high real-time performance. Hence, we proposed a novel convolutional neural network(CNN) method to directly extract features from point cloud for 3D object classification. We firstly train a pre-training model with ModelNet40 dataset. Then, we freeze the first five layers of our CNN model and adjust the learning rate to fine tune our CNN model. Finally, we evaluate our methods by ModelNet40 and the classification accuracy of our model can achieve 87.8% which is better than other traditional approaches. We also design some experiments to research the effect of T-Net proposed by Charles R. Qi et al. on 3D object classification. In the end, we find that T-Net has little effect on classification task and it is not necessary to apply in our CNN.
KW - 3D object classification
KW - Convolution neural network
KW - Point cloud processing
UR - https://www.scopus.com/pages/publications/85130401794
U2 - 10.1007/978-3-031-04245-4_29
DO - 10.1007/978-3-031-04245-4_29
M3 - 会议稿件
AN - SCOPUS:85130401794
SN - 9783031042447
T3 - Lecture Notes of the Institute for Computer Sciences, Social-Informatics and Telecommunications Engineering, LNICST
SP - 333
EP - 344
BT - 6GN for Future Wireless Networks - 4th EAI International Conference, 6GN 2021, Proceedings
A2 - Shi, Shuo
A2 - Ma, Ruofei
A2 - Lu, Weidang
PB - Springer Science and Business Media Deutschland GmbH
T2 - 4th EAI International Conference on 6G for Future Wireless Networks, 6GN 2021
Y2 - 30 October 2021 through 31 October 2021
ER -