Semantic segmentation of building facades from 3D mesh data is an important approach to understanding and analyzing 3D scenes. Most current semantic segmentation methods rely heavily on fully supervised models, which, in turn, depend on a large amount of finely annotated data, making them costly. In response to this issue, a semi-supervised learning approach is proposed, introducing a semi-supervised semantic segmentation method based on contrastive learning (SemiSeg-Contrastive and Consistency regularization, SS_CC) to segment building facades in 3D mesh data. The contrastive learning module in SS_CC utilizes the class discriminability of pixel-level features between positive and negative samples, enabling better utilization of class feature information. Additionally, SS_CC includes a feature space-based consistency regularization loss module that further analyzes the overall features of building facades from a global feature perspective. The experimental results show that the proposed SS-CC method outperforms some mainstream methods in F1 score and mIoU, and has relatively better segmentation performance on building walls and windows.