Gaussian Process Domain Experts for Modeling of Facial Affect

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Facial behavior analysis models struggle with generalization due to variations in subject morphology and camera views.
• Existing generic classifiers perform poorly on unseen data from different contexts.
• Contextual differences between training and testing data significantly impact model performance.

Purpose of the Study:

• To propose a data-efficient domain adaptation approach for adapting existing facial behavior analysis models to new, unseen contexts.
• To leverage Gaussian processes (GPs) to create domain-specific experts for improved model adaptation.
• To enhance model generalization by probabilistically conditioning target experts on source expert predictions.

Main Methods:

• Utilized a probabilistic framework based on Gaussian processes (GPs).
• Introduced domain-specific GP experts, with one expert per subject.
• Facilitated model adaptation by conditioning the target expert on predictions from multiple source experts.
• Employed predictive variance of each expert for optimal weighting during inference.

Main Results:

• The proposed domain adaptation method consistently outperformed both source and target classifiers.
• Achieved superior performance even with a small number of target examples during adaptation.
• Outperformed related state-of-the-art supervised domain adaptation approaches.
• Successfully adapted models for multi-class and multi-label facial expression analysis across different datasets (MultiPIE, DISFA, FERA2015).

Conclusions:

• The proposed Gaussian process-based domain adaptation method effectively improves facial behavior analysis model generalization.
• The approach is data-efficient, requiring minimal target data for adaptation.
• This probabilistic method offers a robust solution for handling contextual variations in facial analysis tasks.