Stsd：A Large-Scale Benchmark for Semantic Segmentation of Subway Tunnel Point Cloud

Laser scanning technology has demonstrated its unique advantages in applications related to subway tunnel scenes, such as health inspection and BIM. Deep learning (DL) semantic segmentation of tunnel point cloud shows an efficient path for these applications and has become a hot topic in recent years. However, these methods often suffer from a shortage of benchmarks and training data on tunnel point cloud segmentation. In this work, a large-scale dataset for semantic segmentation of subway tunnel point cloud was proposed, called subway tunnel segmentation dataset (STSD), which consists of the point cloud of more than 2700 meters of subway tunnels and 2.26 billion points. The STSD contains both point cloud and projected images and is carefully labeled into 11 classes. A novel approach for DL semantic segmentation in subway tunnel point clouds is also proposed. The 3D point cloud of a subway tunnel is converted into a 2.5D point cloud and then projected onto different image channels. With this approach, advanced DL image segmentation networks can be used in subway tunnel point clouds. Previous projection-based methods are only suitable for circular tunnels and tend to convert 3D point clouds directly into intensity images. In comparison, this approach can convert point cloud of subway tunnel with arbitrary curvature or cross-sectional shape into images without any occlusion or scale distortion. Most of the point cloud information, including intensity and geometry, can be preserved when converting to images. In addition, the labeling process can be performed on the 2.5D point cloud, which has significant advantages in efficiency and accuracy over labeling on the original point cloud or projected images. Further evaluation of several classic or state-of-the-art 2D and 3D DL semantic segmentation models shows the feasibility of the approach and dataset. The best 2D model achieves a mIoU of 90.56% and outperforms the best 3D model by almost 10%. This research provides a novel approach for DL semantic segmentation in subway tunnel point clouds, contributes a large-scale dataset for the tunnel inspection domain, and creates a benchmark for further evaluation of the corresponding algorithms.