3D Scene Understanding for Highway Tunnel Construction: Challenges and Baseline Analysis

3D scene understanding is revolutionising tunnel engineering. However, deep learning algorithms are data-hungry, which means the application of scene understanding on tunnel engineering requires a customized point cloud dataset in the construction field. In this paper, we introduce a new point cloud dataset called HTunnel-HLS, specifically designed for construction highway tunnel environment. HTunnel-HLS aims to establish a new database for developing semantic segmentation, and importantly, construction highway tunnel scene. Besides, the dataset provides both point-level semantic labelling along with a large range of types of semantic instance labels categorized into support structures, mechanical facilities, and others. Data have been acquired by the Hand Laser Scanning (HLS) system Hovermap and contains 28 scenes, over 1.58 billion 3D points, correspond to a 9 km long tunnel section. This paper also provides the performance of several representative baseline methods. The impact of scale on model performance is analyzed from the perspective of grid size, and outlines potential future works and challenges for fully exploiting this dataset.