Synthesising Facial Macro- and Micro-Expressions Using Reference Guided Style Transfer.

Area of Science:

• Computer Vision
• Machine Learning
• Biometrics

Background:

• Deep learning methods require extensive long video datasets for facial macro- and micro-expressions.
• Existing methods for generating long micro-expression videos are limited.
• There is a lack of standardized metrics for evaluating synthetic facial expression datasets.

Purpose of the Study:

• To introduce a novel approach for generating synthetic long facial expression videos.
• To propose assessment methods for evaluating the quality of generated datasets.
• To address the scarcity of data for micro-expression research.

Main Methods:

• Utilized StarGANv2, a state-of-the-art generative adversarial network style transfer method.
• Transferred facial expression styles from the SAMM long video dataset onto FFHQ images to create the SAMM-SYNTH dataset.
• Evaluated dataset quality using facial action unit (AU) detection via OpenFace and optical flow analysis.

Main Results:

• Generated synthetic dataset (SAMM-SYNTH) demonstrated high correlation with original data for key Action Units (AU12: 0.74, AU6: 0.72).
• OpenFace evaluation yielded high scores for AU12 (0.85) and AU6 (0.59).
• Optical flow analysis confirmed visual similarity between original and transferred facial movements.

Conclusions:

• The proposed method effectively generates synthetic long videos of facial micro-expressions.
• The SAMM-SYNTH dataset and evaluation metrics contribute to advancing micro-expression research.
• Published dataset aims to increase data availability for micro-expression spotting tasks.