A surface graph based deep learning framework for large-scale urban mesh semantic segmentation

The acquisition of large-scale 3D urban scene by photogrammetry and remote sensing is becoming faster and easier in recent years. As one of the important steps to help machines understand scenarios, mesh semantic segmentation has received extensive attention. Aiming at the 3D urban scene, a surface graph based deep learning framework is proposed, which combines the merits of simple representation of point cloud and ex-presses complex surface topography and texture of mesh. The proposed model employs COG graph to represent surface topography of the mesh. Then novel mesh abstraction and neighborhood definition are conducted on the COG graph. In addition, we propose a texture convolution to extract textual features for individual facets. A hierarchical network architecture is adopted on the prebuilt abstraction and neighborhood data. The experiments on the self-made Wuhan dataset verify the effectiveness of the introduction of surface topography and texture convolution. Additionally, our model increases performance to 94.1% (OA), 71.5% (mIoU) and 79.4% (mF1) in comparative experiments on the SUM dataset that proves its strong competitiveness in semantic segmentation of urban scenes.