Neural marching cubes

We introduce Neural Marching Cubes , a data-driven approach for extracting triangle mesh from discretized implicit field. base our meshing on (MC), due to the simplicity of its input, namely uniform grid signed distances or occupancies, which frequently arise in surface reconstruction and neural models. However, classical MC is defined by coarse tessellation templates isolated individual cubes. While more refined tessellations have been proposed several variants, they all make heuristic assumptions, such as trilinearity, when determining vertex positions local topologies each cube. In principle, none these approaches can reconstruct geometric features that reveal coherence dependencies between nearby cubes (e.g., sharp edge ), information unaccounted for, resulting poor estimates true underlying To tackle challenges, we re-cast deep learning perspective, designing apt at preserving features, training meshes, account contextual develop compact per-cube parameterization represent output mesh, while being compatible with processing, so simple 3D convolutional network be employed training. show topological cases cube are applicable design easily derived using representation, also obtained naturally efficiently following few guidelines. addition, learns limited receptive fields, hence it generalizes well new shapes datasets. evaluate quantitative qualitative comparisons well-known variants. particular, demonstrate ability recover edges corners, long-standing issue Our reconstructs accurately than previous approaches. Code data available https://github.com/czq142857/NMC.