Abstract | ||
---|---|---|
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. |
Year | DOI | Venue |
---|---|---|
2022 | 10.1007/978-3-031-06794-5_31 | Artificial Intelligence and Security |
Keywords | DocType | ISSN |
Point cloud, Semantic segmentation, Multi-task learning | Conference | 0302-9743 |
Citations | PageRank | References |
0 | 0.34 | 0 |
Authors | ||
5 |
Name | Order | Citations | PageRank |
---|---|---|---|
Lin Xi | 1 | 0 | 0.34 |
Luo Huan | 2 | 0 | 0.34 |
Guo Wenzhong | 3 | 0 | 0.34 |
Cheng Wang | 4 | 118 | 29.56 |
Jonathan Li | 5 | 798 | 119.18 |