Geometry Sampling-Based Adaption to DCGAN for 3D Face Generation.

Area of Science:

• Computer Vision
• 3D Reconstruction
• Deep Learning

Background:

• 3D shape acquisition remains a challenge for 3D face applications.
• Existing 2D deep learning models are not directly applicable to 3D data due to dimensionality differences.

Purpose of the Study:

• To propose novel sampling methods for representing 3D faces as structured data suitable for deep networks.
• To enable unsupervised generative 3D face models capable of producing varied expressions.

Main Methods:

• Geometric sampling using iso-geodesic and radial curves for structured 3D face representation.
• Depth-like map sampling using average grid cell depth for structured 3D face representation.
• Adaptation of structured 3D faces to Deep Convolution Generative Adversarial Network (DCGAN) for generation.

Main Results:

• Successfully represented unstructured 3D faces as matrix-like data.
• Enabled the use of DCGAN for unsupervised 3D face generation.
• Demonstrated the generation of diverse 3D faces with different expressions.

Conclusions:

• The proposed sampling methods effectively bridge the gap between 3D face data and deep learning models.
• The generative model can produce a wide variety of 3D faces with controllable expressions.
• This work advances 3D face generation capabilities using deep learning.