One-Frame Calibration Boosts Facial AU Recognition
A new method for automatic facial action unit (AU) recognition uses a single neutral-expression image per face for calibration. Researchers propose one-frame calibration (OFC) to address the infeasibility of accurate AU inference from single images of unseen faces due to facial attribute diversity. They develop a Calibrating Siamese Network (CSN) with a simple iResNet-50 backbone, demonstrating effectiveness on DISFA, DISFA+, and U datasets. The approach contrasts with existing cross-participant non-calibrated generalization (NCG) systems.
Key facts
- One-frame calibration (OFC) uses a single neutral-expression image as reference.
- Calibrating Siamese Network (CSN) is developed for AU recognition.
- CSN uses a simple iResNet-50 backbone.
- Method tested on DISFA, DISFA+, and U datasets.
- Addresses diversity of facial attributes across identities.
- Existing AU systems aim for cross-participant non-calibrated generalization (NCG).
- Accurate AU activation inference from single images of unseen faces is sometimes infeasible.
- Neutral expression understanding is crucial to avoid bias.
Entities
—