Spontaneous Expression Recognition using Universal Attribute Model.

Area of Science:

• Computer Science
• Artificial Intelligence
• Human-Computer Interaction

Background:

• Spontaneous expression recognition is crucial for understanding human emotions.
• Current methods rely on time-consuming manual annotations, limiting scalability.
• A need exists for unsupervised approaches to recognize non-posed facial expressions.

Purpose of the Study:

• To develop an unsupervised framework for spontaneous expression recognition.
• To eliminate the reliance on expert-annotated data.
• To create a robust method for identifying facial expressions in videos.

Main Methods:

• Trained a Universal Attribute Model (UAM) using Gaussian Mixture Models to implicitly learn expression attributes.
• Developed a Super Expression-Vector (SEV) by adapting UAM means for each video clip.
• Utilized factor analysis to decompose SEV into a low-dimensional expression-vector, preserving discriminative information.

Main Results:

• The proposed unsupervised framework achieved superior performance on spontaneous expression datasets (BP4D, AFEW) compared to state-of-the-art techniques.
• The method effectively handles high inter-actor variability in spontaneous expressions.
• A UAM trained on constrained data demonstrated effectiveness for unconstrained expression recognition.

Conclusions:

• The unsupervised framework provides an efficient and effective solution for spontaneous expression recognition.
• The developed expression-vectors are distinct and robust to variations between individuals.
• This approach significantly advances the field by removing the bottleneck of manual data annotation.