Facial Action Unit Detection using 3D Face Landmarks for Pain Detection.

Area of Science:

• Computer Vision
• Biomedical Engineering
• Human-Computer Interaction

Background:

• Manual annotation of facial action units (AUs) is time-consuming and prone to errors.
• Accurate AU detection is crucial for applications like facial expression analysis and pain detection.
• Existing methods for AU detection can be resource-intensive.

Purpose of the Study:

• To develop an efficient method for automatic facial action unit (AU) detection using 3D face landmarks.
• To validate the effectiveness of the AU detection model for pain detection.
• To establish a new benchmark for pain detection using AU analysis.

Main Methods:

• Utilized 3D face landmarks for efficient AU detection.
• Trained deep learning models, including a Transformer model, for pain detection based on predicted AUs.
• Evaluated performance on the BP4D+ dataset.

Main Results:

• The proposed AU detection method is efficient and effective for downstream tasks like pain detection.
• Achieved an 11.13% improvement in F1-score and 3.09% improvement in accuracy for pain detection on the BP4D+ dataset using a Transformer model.
• Demonstrated competitive pain detection results using only eight predicted AUs compared to using 34 ground-truth AUs.

Conclusions:

• Efficient AU detection using 3D landmarks significantly enhances the reliability and speed of facial expression analysis.
• The developed AU detection method provides a robust foundation for accurate pain detection.
• This approach offers a promising direction for advancing automated facial analysis in research and clinical settings.