Mesh Denoising With Facet Graph Convolutions.

Area of Science:

• Computer Graphics
• Geometric Modeling
• Machine Learning

Background:

• Traditional mesh denoising often requires extensive parameter tuning for specific features and noise.
• Existing data-driven methods show promise but may still rely on hand-crafted features or local processing.

Purpose of the Study:

• To develop an end-to-end learning strategy for robust 3D mesh denoising.
• To automatically learn meaningful geometric features directly from mesh data.
• To improve upon current state-of-the-art mesh denoising techniques.

Main Methods:

• A fully end-to-end learning strategy based on graph convolutions operating on the mesh's facet topology.
• A multi-scale design to extract geometric features at various resolution levels.
• Denoising face normals first, followed by vertex position updates.

Main Results:

• The proposed method significantly outperforms current state-of-the-art learning-based mesh denoising approaches.
• The method can be trained effectively using noisy data without explicit ground-truth correspondences.
• A multi-scale strategy is introduced, proving effective for low spatial frequency noise correction.

Conclusions:

• The graph convolution-based approach offers a robust and effective solution for 3D mesh denoising.
• Automatic feature learning eliminates the need for manual parameter fine-tuning.
• The method advances the field by enabling training without paired noisy/clean data and handling multi-scale noise.