Geometry Guided Pose-invariant Facial Expression Recognition.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Facial Expression Recognition (FER) is crucial for human-centered computing applications.
• Conventional FER methods struggle with pose variations, often requiring face frontalization or pose-specific classifiers.

Purpose of the Study:

• To develop an end-to-end deep learning model for simultaneous facial image synthesis and pose-invariant FER.
• To overcome limitations of existing FER approaches by leveraging facial shape geometry.

Main Methods:

• Proposed an end-to-end deep learning model based on Generative Adversarial Networks (GAN).
• Utilized facial landmarks and shape geometry to disentangle identity from expression and pose variations.
• Enabled identity-preserving face generation guided by target pose and expression.

Main Results:

• The model successfully generates identity-preserving faces with specified expressions and poses.
• Explicit disentanglement of identity, expression, and pose variations was achieved.
• Generated diverse face images to augment training datasets for FER.
• Demonstrated superior performance compared to state-of-the-art methods on benchmark datasets (Multi-PIE, BU-3DFE, SFEW).

Conclusions:

• The proposed GAN-based model effectively addresses pose variations in FER.
• Leveraging shape geometry provides a robust method for disentangling facial attributes.
• The approach enhances FER accuracy and dataset augmentation capabilities.