RTF2Mesh: Restricted Tangent Face Based Mesh Compression With Neural Displacement Fields

Area of Science:

• Computer Graphics
• Geometric Modeling
• Neural Networks

Background:

• Encoding 3D mesh surfaces into compact neural representations is a key research area.
• Traditional methods use a base mesh and displacement field, facing limitations in memory overhead and base mesh optimization.
• Explicit storage of vertex connectivity in traditional methods leads to substantial memory requirements.

Purpose of the Study:

• To propose RTF2Mesh, a method for compact neural representation of mesh surfaces.
• To overcome limitations of traditional mesh decomposition approaches.
• To achieve higher compression ratios and representation accuracy without explicit vertex connectivity.

Main Methods:

• RTF2Mesh utilizes unstructured point clouds with feature vectors and network parameters, eliminating explicit mesh storage.
• A meshless vertex-normal representation derived from Restricted Tangent Face (RTF) is employed.
• Kolmogorov-Arnold Networks (KANs) encode displacement and normals, offering superior parameter efficiency over MLPs.

Main Results:

• RTF2Mesh achieves compact neural representation by using only point clouds and KANs.
• The method eliminates the need for explicit vertex connectivity storage.
• RTF2Mesh demonstrates highly competitive performance compared to state-of-the-art methods at equivalent compression rates.

Conclusions:

• RTF2Mesh offers a more compact and efficient neural representation for 3D surfaces.
• The integration of RTF and KANs addresses limitations in existing mesh encoding techniques.
• This approach enables high-resolution mesh generation with reduced memory footprint.