A platform for research: civil engineering, architecture and urbanism
A platform for research: civil engineering, architecture and urbanism
Semi-supervised learning-based point cloud network for segmentation of 3D tunnel scenes
Abstract Automatic identifying target multi-class objects in tunnel scenes from 3D point clouds is widely thought to be critical for maintaining the healthy condition of the tunnel using deep learning methods. However, those methods require extensive data with labels, which is time-consuming and labor-intensive. Targeting effective multi-class tunnel point cloud segmentation for practical applications, this research proposes a deep learning method named semi-supervised learning-based point cloud network (SPCNet) to boost segmentation by alleviating labeling tasks. It contains a supervised learning module, a self-training module, a mean teacher-based learning module, loss functions, and evaluation metrics. To validate the effectiveness and reliability of the proposed method SPCNet, a point cloud collected from a real tunnel is implemented. The results indicate that the proposed method SPCNet performs excellently with MIoU of 0.8741 and 0.8583 in Scenarios I and II, respectively; as well as superior to the supervised learning method with MIoU of 0.8152 and 0.7552 in Scenarios I and II and other state-of-the-art methods such as ST++ and ST. Accordingly, the proposed method SPCNet has superior performance, beneficially contributing to multi-class object segmentation of 3D tunnel point clouds with great potential for applications in practice.
Highlights A deep learning method named SPCNet is developed for 3D point cloud segmentation. It applies unlabeled point clouds to boost segmentation with alleviating labeling tasks. A novel loss function is presented to strengthen feature learning and handle data imbalance. The effectiveness is verified in a dataset with six classes over 32 million tunnel points. The developed method performs excellently with MIoU of 0.8741 and 0.8583 in scenarios.
Semi-supervised learning-based point cloud network for segmentation of 3D tunnel scenes
Abstract Automatic identifying target multi-class objects in tunnel scenes from 3D point clouds is widely thought to be critical for maintaining the healthy condition of the tunnel using deep learning methods. However, those methods require extensive data with labels, which is time-consuming and labor-intensive. Targeting effective multi-class tunnel point cloud segmentation for practical applications, this research proposes a deep learning method named semi-supervised learning-based point cloud network (SPCNet) to boost segmentation by alleviating labeling tasks. It contains a supervised learning module, a self-training module, a mean teacher-based learning module, loss functions, and evaluation metrics. To validate the effectiveness and reliability of the proposed method SPCNet, a point cloud collected from a real tunnel is implemented. The results indicate that the proposed method SPCNet performs excellently with MIoU of 0.8741 and 0.8583 in Scenarios I and II, respectively; as well as superior to the supervised learning method with MIoU of 0.8152 and 0.7552 in Scenarios I and II and other state-of-the-art methods such as ST++ and ST. Accordingly, the proposed method SPCNet has superior performance, beneficially contributing to multi-class object segmentation of 3D tunnel point clouds with great potential for applications in practice.
Highlights A deep learning method named SPCNet is developed for 3D point cloud segmentation. It applies unlabeled point clouds to boost segmentation with alleviating labeling tasks. A novel loss function is presented to strengthen feature learning and handle data imbalance. The effectiveness is verified in a dataset with six classes over 32 million tunnel points. The developed method performs excellently with MIoU of 0.8741 and 0.8583 in scenarios.
Semi-supervised learning-based point cloud network for segmentation of 3D tunnel scenes
Ji, Ankang (author) / Zhou, Yunxiang (author) / Zhang, Limao (author) / Tiong, Robert L.K. (author) / Xue, Xiaolong (author)
2022-11-08
Article (Journal)
Electronic Resource
English
Semi-supervised learning network for deep-sea nodule mineral image segmentation
| Elsevier | 2025
Instance segmentation using semi-supervised learning for fire recognition
| Elsevier | 2022