Abstract: In photogrammetry and remote sensing community, 3D mesh is one of the final user products, which is widely applied in urban planning, navigation, etc. However, there are few works on semantic complex 3D mesh urban scene segmentation based on deep learning methods. Thus, a semantic segmentation method of 3D scenes using dynamic graph CNN for complex city (3Dcity-net) is proposed. By using mesh-inherent features containing 3D spatial information and texture information, a composite feature vector is proposed to represent each face in 3D mesh. To reduce the influence on semantic segmentation by the noise and redundant information in texture information, a principal component analysis (PCA) module is embedded in to the proposed 3D city-net. In order to alleviate the problem of semantic segmentation precision decrease caused by the unbalanced sample data, the focal loss function is used to replace the cross-entropy loss function. The Hessigheim 3D mesh data are utilized to perform experiments. The results of experiments show that the proposed method can obtain competitive semantic segmentation results on 3D mesh. The overall accuracy, Kappa coefficient, mean precision, mean recall, mean F₁ score, and mean IoU is 81.5%, 0.776, 73.0%, 58.4%, 62.6%, and 49.8%, respectively. Comparing to two state-of-the-art methods, the overall accuracy increases by 0.9%, and 8.3%, respectively.

Key words: 3D real scene, semantic segmentation, graph CNN network, 3D representation, 3D mesh