ResGEM: Multi-Scale Graph Embedding Network for Residual Mesh Denoising.

Area of Science:

• Computer Graphics
• Geometric Processing
• Machine Learning

Background:

• 3D mesh denoising is vital for recovering clean geometry from noisy data.
• Deep learning, especially graph convolutional networks (GCNs), shows promise but struggles with irregular topologies.
• Faithfully reconstructing normals and vertices on complex meshes remains a significant challenge.

Purpose of the Study:

• To develop a novel mesh denoising pipeline addressing challenges with irregular topologies.
• To introduce ResGEM, a GCN designed for accurate normal and vertex regression.
• To improve the balance between geometric detail preservation and surface smoothness in denoised meshes.

Main Methods:

• A parallel normal-aware and vertex-aware network architecture.
• ResGEM GCN with multi-scale edge-conditioned embedding modules (EEMs) for feature extraction.
• Residual learning framework for predicting normal and vertex offsets, incorporating novel regularization terms.

Main Results:

• The proposed method effectively denoises 3D meshes, preserving geometric details and surface topology.
• ResGEM demonstrates superior performance compared to state-of-the-art methods on synthetic and real-scanned datasets.
• Novel regularization terms enhance network smoothing and generalization capabilities.

Conclusions:

• The novel pipeline and ResGEM GCN offer a robust solution for 3D mesh denoising, especially for meshes with irregular topology.
• The approach successfully balances geometric fidelity and surface smoothness.
• Experimental results validate the method's superiority over existing techniques.