In heavily occluded underground scenarios such as coal mine roadways, high-precision 3D perception plays an essential role in infrastructure inspection, disaster prevention and control. However, existing general-purpose point cloud semantic segmentation methods suffer from poor generalization robustness in underground environments, due to challenges including blurred boundaries, strong noise interference, and highly repetitive structural patterns. They also typically incur excessive computational overhead, limiting their practical deployment. To address these issues, we propose an efficient and robust point cloud semantic segmentation framework tailored for complex underground infrastructure. Specifically, we design a lightweight Group-aware Channel Interaction Module (GCIM) with weight-shared grouped channel attention and explicit inter-group communication, which enlarges the receptive field for complex scene understanding while maintaining low computation costs. Furthermore, we present Interactive Position Encoding (IPE), which dynamically couples point features with spatial context via relative positional information. By strengthening spatial-aware feature interaction, IPE effectively mitigates boundary ambiguity caused by noise and geometric similarity. Experimental results on underground datasets including Coal Mines, Seg2Tunnel and OpenTrench3D, as well as general indoor datasets ScanNetv2 and S3DIS, demonstrate that the proposed method achieves superior performance in terms of mean Intersection over Union and Overall Accuracy. It exhibits outstanding segmentation capability and cross-scene generalization ability, which can provide effective technical support for intelligent 3D information extraction and safety monitoring in underground environments.
Liu et al. (Mon,) studied this question.