Learning Self-Prior for Mesh Inpainting Using Self-Supervised Graph Convolutional Networks.

Area of Science:

• Computer Graphics
• Geometric Modeling
• Artificial Intelligence

Background:

• 3D mesh completion is crucial for various applications.
• Existing methods often require large datasets or convert meshes to intermediate formats.
• Dataset-independent approaches are needed for robustness and handling novel shapes.

Purpose of the Study:

• To present a self-prior-based mesh inpainting framework.
• To avoid reliance on training datasets and intermediate shape representations.
• To maintain the polygonal mesh format throughout the inpainting process.

Main Methods:

• Introduced single-resolution GCN (SGCN) and multi-resolution GCN (MGCN).
• Employed a self-supervised training strategy using 'fake holes'.
• Refined an initial watertight mesh via vertex displacements predicted by GCNs.

Main Results:

• The framework successfully inpaints incomplete meshes without training data.
• Maintained the polygonal mesh format, avoiding voxel grids or point clouds.
• Outperformed traditional dataset-independent methods and showed robustness against deep-learning alternatives for rare shapes.

Conclusions:

• The proposed GCN-based mesh inpainting framework offers a dataset-free and robust solution.
• Preserving the mesh format simplifies the process and enhances applicability.
• The self-supervised approach with fake holes enables effective learning of vertex displacements for accurate completion.