Explicit Facial Expression Transfer via Fine-Grained Representations.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Facial expression transfer is complex due to intertwined attributes.
• Existing methods rely on potentially inaccurate intermediate predictions (landmarks, AUs).

Purpose of the Study:

• To develop a novel method for explicit facial expression transfer between unpaired images.
• To overcome limitations of indirect expression manipulation methods.

Main Methods:

• Propose a multi-class adversarial training to disentangle images into AU-related and AU-free features.
• Synthesize new images by combining AU-free features with swapped AU-related features.
• Introduce a swap consistency loss for reliable unpaired expression transfer.

Main Results:

• The proposed method achieves state-of-the-art performance in fine-grained facial expression transfer.
• Successfully preserves identity and pose attributes during expression swapping.
• Demonstrates superior results compared to existing expression manipulation techniques.

Conclusions:

• Directly mapping images and disentangling features offers a more robust approach to expression transfer.
• The novel method effectively handles unpaired image data for expression manipulation.
• This work advances the field of facial attribute manipulation and image synthesis.