Point Cloud-Based Segmentation of Small Roof Components in Chinese Ancient Architecture

Abstract. The roofs of ancient Chinese buildings are rich in cultural symbolism, embodying profound historical and artistic significance. To preserve the structural and semantic information of these roof components, this study employs point cloud semantic segmentation, as point clouds effectively capture their authentic geometry and dimensions. To reduce the high cost of manual annotation, we propose a weakly supervised learning approach for point cloud segmentation. However, a significant challenge arises due to the overwhelming presence of roof tiles in the point cloud data, which hinders segmentation performance. Since tiles constitute the majority of the point cloud, smaller architectural components become underrepresented. As a result, when ground truth labels are assigned randomly, the number of labeled points for these smaller elements is insufficient, leading to suboptimal segmentation accuracy. To address this issue, we refine the positional encoding method based on advancements in the attention mechanism, thereby enhancing the model’s ability to focus on small-scale components. Experimental results demonstrate that our approach achieves a 38.61% improvement in mean Intersection over Union (mIoU) compared to SQN, along with a 3.36% increase in overall accuracy (OA). Notably, our method even outperforms certain fully supervised networks in segmentation effectiveness.