TY - GEN
T1 - A Multi-task Learning Framework for Semantic Segmentation in MLS Point Clouds
AU - Lin, Xi
AU - Luo, Huan
AU - Guo, Wenzhong
AU - Wang, Cheng
AU - Li, Jonathan
N1 - Publisher Copyright:
© 2022, The Author(s), under exclusive license to Springer Nature Switzerland AG.
PY - 2022
Y1 - 2022
N2 - We study a multi-task learning framework for semantic segmentation in Mobile Laser Scanning (MLS) point clouds. The existing methods on semantic segmentation of point cloud rely on a large number of annotation data. However, manually annotating data is time-consuming and laborious, and the manually annotation efficiency is particularly low. To alleviate those problems, we propose to exploit a multi-task learning framework to reduce the large demand of training samples for implementing semantic labeling of point clouds. Specifically, we design a new neural network containing a backbone network and two branching networks, which accomplish the color prediction and category prediction, respectively. Color prediction, as an auxiliary task, can be easily conducted by exploiting the color information of each 3D point to train the proposed neural network. Here, color information of each point can be easily generated by using the optical images obtained by the cameras equipped in the MLS system. Once the training procedure of color prediction is completed, we only use a small portion of manually-annotated points to fine-tune the branching network of category prediction for each 3D point. To demonstrate the effectiveness and correctness of our proposed framework, we conducted extensive experiments on the colorized point clouds which are collected by a RIEGL VMX450 MLS system. The experimental results show the proposed approach can reach 96.04%. OA and 94.41% mIoU under the supervision of 10% annotation data.
AB - We study a multi-task learning framework for semantic segmentation in Mobile Laser Scanning (MLS) point clouds. The existing methods on semantic segmentation of point cloud rely on a large number of annotation data. However, manually annotating data is time-consuming and laborious, and the manually annotation efficiency is particularly low. To alleviate those problems, we propose to exploit a multi-task learning framework to reduce the large demand of training samples for implementing semantic labeling of point clouds. Specifically, we design a new neural network containing a backbone network and two branching networks, which accomplish the color prediction and category prediction, respectively. Color prediction, as an auxiliary task, can be easily conducted by exploiting the color information of each 3D point to train the proposed neural network. Here, color information of each point can be easily generated by using the optical images obtained by the cameras equipped in the MLS system. Once the training procedure of color prediction is completed, we only use a small portion of manually-annotated points to fine-tune the branching network of category prediction for each 3D point. To demonstrate the effectiveness and correctness of our proposed framework, we conducted extensive experiments on the colorized point clouds which are collected by a RIEGL VMX450 MLS system. The experimental results show the proposed approach can reach 96.04%. OA and 94.41% mIoU under the supervision of 10% annotation data.
KW - Multi-task learning
KW - Point cloud
KW - Semantic segmentation
UR - https://www.scopus.com/pages/publications/85135065831
U2 - 10.1007/978-3-031-06794-5_31
DO - 10.1007/978-3-031-06794-5_31
M3 - 会议稿件
AN - SCOPUS:85135065831
SN - 9783031067938
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 382
EP - 392
BT - Artificial Intelligence and Security - 8th International Conference, ICAIS 2022, Proceedings
A2 - Sun, Xingming
A2 - Zhang, Xiaorui
A2 - Xia, Zhihua
A2 - Bertino, Elisa
PB - Springer Science and Business Media Deutschland GmbH
T2 - 8th International Conference on Artificial Intelligence and Security, ICAIS 2022
Y2 - 15 July 2022 through 20 July 2022
ER -