Facial expression recognition via variational inference.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Psychology

Background:

• Facial expressions in real-world scenarios are complex, often appearing as compound emotions or subtle variations.
• Conventional facial expression recognition (FER) models struggle with this complexity due to deterministic estimation, failing to capture the continuous and uncertain nature of emotions.

Purpose of the Study:

• To develop a novel framework, POSTER-Var, that addresses the limitations of existing FER methods by modeling the probabilistic nature of facial expressions.
• To enhance feature representation for more effective fusion of hierarchical semantic information in FER.

Main Methods:

• Proposed POSTER-Var framework incorporating a Variational Inference-based Classification Head (VICH).
• VICH utilizes the reparameterization trick to map facial features into a probabilistic latent space, learning expression intensity distributions.
• Introduced layer embeddings and nonlinear transformations within a feature pyramid to improve feature representation.

Main Results:

• POSTER-Var demonstrated superior performance in handling fine-grained expression recognition.
• Achieved state-of-the-art results on benchmark datasets including RAF-DB, AffectNet, and FER+.
• The framework effectively models the uncertainty and continuous nature of facial expressions.

Conclusions:

• The proposed POSTER-Var framework with VICH significantly advances the field of facial expression recognition.
• This probabilistic approach offers a more robust solution for analyzing complex and subtle facial expressions in unconstrained environments.
• Open-sourcing the code facilitates further research and development in FER.