POCE: Pose-Controllable Expression Editing.

Area of Science:

• Computer Vision
• Deep Learning
• Image Synthesis

Background:

• Facial expression editing is advancing with deep neural networks.
• Existing methods struggle with pose variations and require paired data.
• This limits realistic editing and usability.

Purpose of the Study:

• To present POCE, a pose-controllable expression editing network.
• To achieve realistic facial expression and head pose generation simultaneously.
• To utilize unpaired training images for accessible editing.

Main Methods:

• POCE maps face images into UV space for disentangled editing.
• Self-supervised UV completion addresses occlusions and missing textures.
• Weakly-supervised UV editing generates new expressions with identity preservation.

Main Results:

• POCE effectively learns from unpaired face images.
• The network generates realistic, high-fidelity facial expressions.
• Editing is controllable under various new head poses.

Conclusions:

• POCE offers an innovative solution for pose-controllable facial expression editing.
• The method enhances realism and usability by disentangling pose and expression.
• It demonstrates strong performance with unpaired data.