Crossing Domains for AU Coding: Perspectives, Approaches, and Measures.

Area of Science:

• Computer Vision
• Human Behavior Analysis
• Machine Learning

Background:

• Facial action unit (AU) detection models excel within trained domains.
• Cross-domain generalizability of these detectors remains a significant challenge.

Purpose of the Study:

• To evaluate the performance of AU detectors across diverse, unseen datasets.
• To identify factors influencing cross-domain transferability in AU detection.

Main Methods:

• Literature review on cross-domain transfer learning for AU detection.
• Experimental evaluation using four distinct public databases (EB+, Sayette GFT, DISFA, UNBC-SP).
• Comparison of deep and shallow learning approaches for AU detection generalizability.

Main Results:

• AU detector performance significantly decreased across different domains, often below research thresholds.
• Deep learning models showed slightly better average performance than shallow models.
• Occlusion sensitivity maps indicated reduced local specificity in cross-domain detection.

Conclusions:

• Current AU detectors lack robust generalizability across varied observational contexts.
• Future research should focus on diverse datasets and advanced deep learning to enhance transferability.
• Caution is advised when deploying AU classifiers in new domains without validation.