Robust face alignment under occlusion via regional predictive power estimation.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Face alignment models struggle with heavy partial occlusion.
• Existing methods often require explicit occlusion modeling or visibility annotations.

Purpose of the Study:

• To improve face alignment performance on occluded images without explicit occlusion modeling.
• To introduce a model adaptation scheme using local regression forest (RF) voting results.

Main Methods:

• A model adaptation scheme is proposed, utilizing regional predictive power (RPP) derived from local RF votes.
• The RPP of oversegmented regions determines the reliability of facial landmark predictions.
• A holistic voting method (cascaded pose regression) is adapted by weighting votes based on RPP.

Main Results:

• The proposed method demonstrates superior performance on challenging datasets (COFW, 300-W) compared to existing face alignment models.
• Achieved high accuracy (72.4% overlap ratio) in estimating face/non-face regions on occluded images without explicit occlusion modeling.

Conclusions:

• The proposed model adaptation scheme effectively addresses performance degradation in face alignment under heavy occlusion.
• This approach offers a robust solution for occluded face alignment and object segmentation without requiring explicit occlusion annotations.