Neural dual contouring

We introduce neural dual contouring (NDC), a new data-driven approach to mesh reconstruction based on (DC). Like traditional DC, it produces exactly one vertex per grid cell and quad for each edge intersection, natural efficient structure reproducing sharp features. However, rather than computing locations crossings with hand-crafted functions that depend directly difficult-to-obtain surface gradients, NDC uses network predict them. As result, can be trained produce meshes from signed or unsigned distance fields, binary voxel grids, point clouds (with without normals); open surfaces in cases where the input represents sheet partial surface. During experiments five prominent datasets, we find NDC, when of generalizes well others. Furthermore, provides better accuracy, feature preservation, output complexity, triangle quality, inference time comparison previous learned (e.g., marching cubes, convolutional occupancy networks) Poisson) methods. Code data are available at https://github.com/czq142857/NDC.