Abstract
We introduce TetWeave, a novel isosurface representation for gradient-basedmesh optimization that jointly optimizes the placement of a tetrahedral gridused for Marching Tetrahedra and a novel directional signed distance at eachpoint. TetWeave constructs tetrahedral grids on-the-fly via Delaunaytriangulation, enabling increased flexibility compared to predefined grids. Theextracted meshes are guaranteed to be watertight, two-manifold andintersection-free. The flexibility of TetWeave enables a resampling strategythat places new points where reconstruction error is high and allows toencourage mesh fairness without compromising on reconstruction error. Thisleads to high-quality, adaptive meshes that require minimal memory usage andfew parameters to optimize. Consequently, TetWeave exhibits near-linear memoryscaling relative to the vertex count of the output mesh - a substantialimprovement over predefined grids. We demonstrate the applicability of TetWeaveto a broad range of challenging tasks in computer graphics and vision, such asmulti-view 3D reconstruction, mesh compression and geometric texturegeneration.