One-shot many-to-many facial reenactment using Bi-Layer Graph Convolutional Networks

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Facial reenactment animates source faces using driving images.
• Existing methods struggle with few-shot scenarios and identity protection.
• Previous research often requires multiple images per identity.

Purpose of the Study:

• To introduce a novel one-shot many-to-many facial reenactment model.
• To address limitations of current facial reenactment techniques in few-shot settings.
• To enable high-quality facial reenactment from a single source image.

Main Methods:

• Developed a Bi-Layer Graph Convolutional Layers (BGCLN) model.
• Utilized a bi-layer decomposition approach with Convolutional Neural Networks (CNN).
• Generated optical flow representation from latent vectors for precise motion simulation.

Main Results:

• Achieved high-quality facial reenactment using only one source image.
• Outperformed recent techniques in both qualitative and quantitative comparisons.
• Demonstrated near real-time performance at 15 frames per second.

Conclusions:

• The BGCLN model offers a significant advancement in one-shot facial reenactment.
• The technique effectively handles identity preservation and motion simulation from minimal data.
• The proposed method provides a robust and efficient solution for facial reenactment tasks.