With the advancement of 3D modeling and deep learning technologies, achieving automated and high-precision annotation on textured 3D triangular mesh models has become a research hotspot. This paper proposes a fine-grained deterioration segmentation mechanism for textured 3D triangular mesh data, based on cross-modal feature extraction and fusion via bidirectional 2D–3D mapping. By constructing a 2D–3D modality mapping model, the method integrates 3D annotation with 2D feature extraction, and finally projects the deterioration mask information back onto the 3D model. Taking a polychrome wooden Guanyin sculpture from the Song Dynasty of China as a case study, the experimental results verify the correctness and feasibility of the proposed method. This automatic annotation approach for textured 3D triangular mesh models fills the gap in automated labeling of 3D data and provides a dataset preparation strategy for subsequent multimodal 3D segmentation using deep learning, demonstrating both scientific value and practical significance.
Hu et al. (Fri,) studied this question.