Reproduce “Mesh Draping: Parametrization-Free Neural Mesh Transfer” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Amir Hertz, Or Perel, Raja Giryes, Olga Sorkine-Hornung, Daniel Cohen-Or
Mesh Draping transfers an artist-designed mesh to a related target while retaining its faces and connectivity. It optimizes vertex offsets to balance target alignment with preservation of structure, using gradually introduced positional frequencies. In a selected ShapeNet comparison, its combined transfer score was 0.933, versus 0.407 for ShapeFlow and 0.361 for Neural Cages; it is not intended for dissimilar shapes.
Key insights
- Progressive Positional Encoding (PPE) enables a neural network to learn both global deformations and fine geometric details in a stable, coarse-to-fine manner, overcoming typical spectral bias issues.
- A parameterization-free neural optimization approach, guided by sparse user correspondences, can effectively transfer intricate mesh structures to diverse target geometries, including non-manifold ones.
- A carefully designed compound loss function, balancing target alignment with structural preservation terms (angle, local area, skinny face), is crucial for maintaining the quality and artistic intent of the source mesh during deformation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.