UMeshSegNet: Semantic Segmentation of 3D Mesh Generated from UAV Photogrammetry

3D mesh generated from UAV photogrammetry can depicts the urban scene realistically. Most of the studies on semantic segmentation of 3D mesh based on deep learning convert mesh data into point cloud or 2D image, resulting in original information lost and poor segmentation effect. To address the problem, a semantic segmentation convolutional neural network UMeshSegNet is designed in this paper based on MeshCNN, which directly processes the mesh data. The network combines geometric, elevation and texture features, and attention mechanism is also introduced to enhance the sensitivity to the feature. Experiments and analyses are conducted on public dataset SUM and our own Wuhan test data, and the experimental results indicate that UMeshSegNet can effectively segment mesh data with significantly higher semantic segmentation accuracy than previous deep learning methods.