Reproduce “Learning Self-Prior for Mesh Inpainting Using Self-Supervised Graph Convolutional Networks” on OpenResearch

About this paper

Shota Hattori, Tatsuya Yatagawa, Yutaka Ohtake, Hiromasa Suzuki

Researchers from The University of Tokyo and Hitotsubashi University developed a self-supervised Graph Convolutional Network (GCN) framework for direct mesh inpainting, leveraging a self-prior from a single incomplete mesh. This method effectively fills holes in 3D triangular meshes without requiring large training datasets or converting mesh formats, demonstrating superior performance over traditional methods and robustness to out-of-distribution shapes.

Key insights

• The network structure itself, when trained on only a single input mesh, acts as a powerful 'self-prior' for inferring missing geometry, circumventing the need for large external datasets.
• Chebyshev spectral graph convolutions (ChebConv) provide more stable training and better performance compared to GCNConv for this specific mesh inpainting task.
• A simple encoder-decoder MGCN (without skip connections) outperformed U-net-like architectures, suggesting that skip connections can hinder the learning of meaningful deformations in self-prior settings, similar to observations in Deep Image Prior.

Summary by alphaXiv.