Cross-Domain Facial Expression Recognition: A Unified Evaluation Benchmark and Adversarial Graph Learning

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Facial expression recognition (FER) has advanced, but dataset inconsistencies limit model generalization.
• Existing cross-domain FER (CD-FER) algorithms lack standardized evaluation, hindering fair comparisons.
• Current CD-FER methods often overlook transferable local features, focusing on holistic domain-invariant features.

Purpose of the Study:

• To establish a unified benchmark for fair and comprehensive evaluation of CD-FER algorithms.
• To introduce a novel framework, adversarial graph representation adaptation (AGRA), for improved cross-domain FER.
• To address the limitations of existing methods by integrating holistic and local feature adaptation.

Main Methods:

• Developed a unified CD-FER evaluation benchmark using consistent datasets and feature extractors.
• Proposed the AGRA framework, combining graph representation propagation with adversarial learning.
• Utilized stacked graph convolution networks (GCNs) for holistic-local feature co-adaptation across domains.

Main Results:

• The unified benchmark enabled fair comparisons of various CD-FER and domain adaptation algorithms.
• The AGRA framework demonstrated superior performance compared to state-of-the-art methods on the benchmark.
• AGRA effectively learns fine-grained, domain-invariant features by adapting both holistic and local information.

Conclusions:

• The proposed unified benchmark facilitates rigorous evaluation in cross-domain facial expression recognition.
• The AGRA framework offers a significant advancement in cross-domain FER by co-adapting holistic and local features.
• This approach enhances model generalization and accuracy across diverse facial expression datasets.