Robust and Precise Facial Landmark Detection by Self-Calibrated Pose Attention Network

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Fully supervised facial landmark detection methods struggle with large poses and occlusions due to inaccurate shape constraints and limited labeled data.
• Existing methods lack robustness in real-world scenarios with significant variations in face orientation and partial visibility.

Purpose of the Study:

• To propose a semisupervised framework, the self-calibrated pose attention network (SCPAN), for robust and precise facial landmark detection.
• To address limitations of current methods in handling large poses and heavy occlusions.
• To improve facial landmark detection accuracy by leveraging unlabeled data.

Main Methods:

• Developed a boundary-aware landmark intensity (BALI) field to enhance facial shape constraints by integrating boundary and landmark intensity information.
• Introduced a self-calibrated pose attention (SCPA) model with a self-calibrated mechanism and pose attention mask for intermediate supervision without label information.
• Integrated BALI fields and SCPA model into the novel SCPAN framework.

Main Results:

• The proposed SCPAN framework effectively learns facial prior knowledge, improving detection accuracy and robustness.
• Experiments on challenging benchmark datasets demonstrate superior performance compared to state-of-the-art methods.
• The method shows significant improvements in detecting facial landmarks under large poses and heavy occlusions.

Conclusions:

• The novel SCPAN framework offers a robust and precise solution for facial landmark detection in challenging scenarios.
• The integration of BALI fields and SCPA model enhances the ability to handle variations in pose and occlusion.
• This semisupervised approach advances the field by effectively utilizing unlabeled data.