Fast-GANFIT: Generative Adversarial Network for High Fidelity 3D Face Reconstruction.

Area of Science:

• Computer Vision
• Computer Graphics
• Machine Learning

Background:

• Deep convolutional neural networks (DCNNs) are widely used for 3D face reconstruction from single images.
• Current methods struggle to reconstruct high-frequency details in facial textures.
• Existing approaches often rely on linear texture spaces or auto-encoders for texture feature learning.

Purpose of the Study:

• To develop a novel approach for reconstructing high-frequency facial texture and shape from single images.
• To improve the photorealism and identity preservation of 3D face reconstructions.
• To overcome limitations of existing methods in capturing fine facial texture details.

Main Methods:

• Utilizing generative adversarial networks (GANs) to train a powerful facial texture prior from a large-scale 3D texture dataset.
• Revisiting 3D Morphable Models (3DMMs) fitting with non-linear optimization for optimal latent parameter estimation.
• Proposing a novel self-supervised regression-based approach for robust initialization and expedited fitting.

Main Results:

• Achieved excellent results in photorealistic and identity-preserving 3D face reconstructions.
• Demonstrated high-frequency facial texture reconstruction, a novel achievement in the field.
• The proposed self-supervised method enhances robustness and speed in the fitting process.

Conclusions:

• The integration of GANs and DCNNs enables superior facial texture and shape reconstruction from single images.
• This approach successfully captures high-frequency details, advancing the state-of-the-art in 3D face modeling.
• The novel self-supervised regression method offers a more robust and efficient fitting process for 3DMMs.