CTWSeg: component-level semantic segmentation of the Ming city wall with a geometry-aware weakly supervised framework

City walls in ancient urban systems served as vital military defenses and symbols of political legitimacy. Although point cloud segmentation has become a key tool in heritage conservation, most studies focus on overall geometric reconstruction and lack component-level semantic understanding of large brick-and-stone structures. We present CTW3D-Parts, a component-level benchmark for Ming city wall point clouds derived from UAV-based multi-view photogrammetry, with seven component categories and standardized annotations. Building on it, we propose CTWSeg, a geometry-aware weakly supervised framework integrating geometric feature encoding, direction-aware structural attention, and confidence-controlled pseudo-label optimization. Using only 0.1% labeled points, CTWSeg achieves 93.14% OA, 86.33% mAcc, and 71.11% mIoU on CTW3D-Parts, ranking highest in OA and mIoU among evaluated methods. Experiments on the ARCH dataset further indicate its robustness. This work provides a benchmark and an effective framework for fine-grained semantic analysis and digital conservation of large-scale architectural heritage.