Action Unit Detection by Learning the Deformation Coefficients of a 3D Morphable Model.

Area of Science:

• Computer Vision
• Computer Graphics
• Biomedical Engineering

Background:

• Facial Action Units (AUs) represent muscle contractions causing facial deformations.
• Existing 3D Morphable Models (3DMMs) struggle with local, asymmetric facial changes.
• Accurate 3D AU analysis and synthesis are crucial for various applications.

Purpose of the Study:

• To develop a novel 3DMM capable of modeling local and asymmetric facial deformations for AUs.
• To enable subject-independent 3D AU detection and synthesis.
• To improve the accuracy and realism of 3D facial expression analysis.

Main Methods:

• A new 3DMM was developed to specifically address local and asymmetric facial deformations.
• Deformation coefficients were learned to decouple identity and expression variations.
• An AU classifier was trained using learned coefficients, and coefficients were applied to neutral scans for AU synthesis.

Main Results:

• The proposed 3DMM demonstrated competitive performance in 3D AU detection on the Bosphorus dataset.
• The approach achieved strong results even in cross-dataset validation settings.
• Learned coefficients proved effective for synthesizing realistic 3D facial instances with activated AUs.

Conclusions:

• The novel 3DMM effectively models complex facial deformations for AUs.
• The method allows for subject-independent 3D AU synthesis, enhancing realism.
• This work advances 3D facial expression analysis and generation capabilities.