Dual hierarchical attention-enhanced transfer learning for semantic segmentation of point clouds in building scene understanding

Targeted to the challenge of indoor scene understanding for intelligent devices, this paper question focuses on enhancing accuracy in semantic information extraction. A framework including a dual hierarchical attention network, transfer learning, interpretability analysis, and modeling module is applied to segment and reconstruct the indoor scene. A high-rise as-built building case is used to verify the method, the results show that: (1) the method achieves a high mIoU of 0.970 in point cloud segmentation and outperforms state-of-the-art methods, both demonstrating strong performance; (2) the method has sound feature extraction and learning ability in term of the interpretive analysis; (3) the method accelerates by 37 % than manual operations, achieving higher accuracy and efficiency. Overall, the method provides an effective solution to segment multi-class objects for indoor scene understanding and can serve as a basis for automated modeling to contribute to an accurate BIM model with great potential for practical application.