GDR-Net: A Geometric Detail Recovering Network for 3D Scanned Objects

Area of Science:

• Computer Graphics
• Geometric Modeling
• Deep Learning

Background:

• 3D mesh super-resolution aims to enhance low-resolution models by recovering intricate geometric details.
• Nonregularity of 3D mesh representations and high complexity of 3D shapes pose significant challenges.
• Existing methods struggle with accurately reconstructing fine geometric features.

Purpose of the Study:

• To propose a novel deep neural network, GDR-Net, for effective 3D mesh super-resolution.
• To address the challenges of nonregularity and complexity in 3D shape representation.
• To generate high-quality 3D models with recovered geometric details.

Main Methods:

• GDR-Net regresses displacements in radial basis function parameter space to handle nonregularity.
• It processes small surface patches to manage high complexity, fusing refined patches for the final mesh.
• A dataset of real-world and synthetic high/low-quality mesh pairs was constructed for training.

Main Results:

• GDR-Net successfully recovers geometric details in general 3D models.
• Experimental results show superior performance compared to previous methods in detail recovery.
• The network effectively enhances the quality of low-resolution 3D meshes.

Conclusions:

• GDR-Net offers a robust solution for the ill-posed problem of mesh super-resolution.
• The proposed approach effectively handles nonregularity and complexity in 3D shapes.
• This method advances the state-of-the-art in generating high-fidelity 3D models from low-resolution inputs.