The key idea is to focus the learning effort of a Neural Unsigned Distance Function (Neural UDF) on surface edges, which improves the fidelity of the obtained Neural UDF to the original 3D surface.
