Synthetic Expressions are Better Than Real for Learning to Detect Facial Actions.

Area of Science:

• Computer Vision
• Machine Learning
• Human-Computer Interaction

Background:

• Training facial action classifiers faces challenges due to small annotated video datasets and infrequent action occurrences.
• Existing methods struggle with data scarcity and imbalanced action frequencies, limiting classifier robustness.

Purpose of the Study:

• To develop and evaluate a novel approach for facial action detection using synthesized facial expression data.
• To overcome limitations of small datasets and low action frequencies in facial expression recognition.

Main Methods:

• Reconstructing 3D facial shapes from video frames and aligning them to a canonical view.
• Employing a Generative Adversarial Network (GAN)-based model to synthesize novel facial expression images.
• Training deep neural networks on both synthesized and real facial expression datasets for comparative analysis.

Main Results:

• The deep neural network trained on synthesized facial expressions significantly outperformed the network trained on unaltered real video.
• The proposed method surpassed current state-of-the-art approaches in facial action detection accuracy.
• Synthesized data effectively augmented limited real-world facial expression datasets.

Conclusions:

• Facial expression generation using GANs offers a promising solution to data scarcity in facial action detection.
• This approach enhances classifier performance and robustness, paving the way for more accurate facial expression recognition systems.
• The study demonstrates the potential of synthetic data generation for advancing machine learning in computer vision tasks.