D-PAttNet: Dynamic Patch-Attentive Deep Network for Action Unit Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Human-Computer Interaction

Background:

• Automated facial action unit (AU) detection is crucial for understanding human emotions and expressions.
• Current methods often struggle with head rotation, predefined AU-patch relationships, and the complex dynamics of AU co-occurrences.
• Existing approaches may ignore the simultaneous nature of AU dynamics as perceived by humans.

Purpose of the Study:

• To develop a robust automated system for facial action unit (AU) detection.
• To overcome limitations of existing AU detection methods, including head rotation and AU co-occurrence.
• To incorporate spatiotemporal dynamics of AUs for more accurate perception.

Main Methods:

• Proposed a dynamic patch-attentive deep network (D-PAttNet) for AU detection.
• Implemented 3D head and face rotation control within the network architecture.
• Developed a method to learn patch-to-AU mappings and model spatiotemporal dynamics simultaneously.

Main Results:

• D-PAttNet significantly improves upon existing state-of-the-art methods for AU detection.
• The model demonstrates enhanced robustness to variations in head pose.
• Simultaneous modeling of spatiotemporal dynamics leads to more accurate AU recognition.

Conclusions:

• The D-PAttNet model offers a significant advancement in automated facial action unit detection.
• The proposed approach effectively handles head rotation and learns complex AU relationships.
• This work contributes to more sophisticated and human-like interpretation of facial expressions.