The preservation of tangible cultural heritage artifacts, particularly those created through intangible cultural heritage (ICH) practices such as traditional craftsmanship, is essential for sustainable urban development. Such artifacts embody the skills and knowledge transmitted through ICH. Digital twin technology offers a systematic approach to document and preserve these artifacts digitally. However, obtaining complete 3D models through close-range photogrammetry remains challenging due to occlusion, surface properties, and sensor limitations, resulting in incomplete point cloud data. This study proposes a generative adversarial network with self-attention mechanism to complete missing 3D point cloud data of cultural heritage artifacts. The network employs a multi-layer perceptron for global feature extraction, self-attention modules for local detail capture, and a feature pyramid decoder for hierarchical point cloud generation. Quantitative evaluation on the ShapeNet dataset demonstrates that the proposed method achieves average errors of 5.541 (P→GT) and 4.183 (T→P) for complete point cloud completion, outperforming FinerPCN, PF-Net, and PFG-Net. For missing point cloud regions, the method achieves errors of 24.303 (P→GT) and 20.008 (GT→P). The proposed framework integrates digital twin concepts with deep learning-based 3D reconstruction to enhance the digital preservation of cultural heritage artifacts. By generating more complete and accurate point clouds, the method enables higher-quality 3D models suitable for virtual exhibition, documentation, and cultural transmission. A limitation is that validation on diverse heritage artifact geometries beyond ShapeNet has not yet been conducted.
Zhang et al. (Thu,) studied this question.