Scalable High-Fidelity 3D Hand Shape Reconstruction via Graph-Image Frequency Mapping and Graph Frequency...

Area of Science:

• Computer Vision
• 3D Graphics
• Machine Learning

Background:

• Current single-image 3D hand modeling techniques struggle to capture intricate mesh details.
• This limitation hinders applications requiring high-fidelity, personalized hand models.

Purpose of the Study:

• To develop a novel network for generating detailed 3D hand meshes from single images.
• To address the limitations of existing methods in capturing high-frequency details for personalized modeling.

Main Methods:

• A frequency split network generating 3D hand meshes in a coarse-to-fine manner across different frequency bands.
• Novel frequency decomposition loss and an image-graph ring feature mapping strategy.
• Bidirectional registration for topology-fixed ground-truth and a new Mean-frequency Signal-to-Noise Ratio (MSNR) metric for evaluation.

Main Results:

• The proposed network successfully preserves high-frequency details, leading to fine-grained 3D hand reconstruction.
• The scalable network adapts to varying computational resources.
• The MSNR metric effectively evaluates the recovery of personalized shape details.

Conclusions:

• The frequency split network significantly enhances the fidelity of 3D hand reconstruction from single images.
• The developed method and evaluation metric advance personalized 3D hand modeling capabilities.