Scalable Surface Reconstruction with Delaunay‐Graph Neural Networks

We introduce a novel learning-based, visibility-aware, surface reconstruction method for large-scale, defect-laden point clouds. Our approach can cope with the scale and variety of cloud defects encountered in real-life Multi-View Stereo (MVS) acquisitions. relies on 3D Delaunay tetrahedralization whose cells are classified as inside or outside by graph neural network an energy model solvable cut. model, making use both local geometric attributes line-of-sight visibility information, is able to learn from small amount synthetic training data generalizes Combining efficiency deep learning methods scalability based models, our outperforms non learning-based algorithms two publicly available benchmarks. code at https://github.com/raphaelsulzer/dgnn.