Topology-Guided Semantic Face Center Estimation for Rotation-Invariant Face Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Face detection accuracy degrades significantly with in-plane (RIP) and out-of-plane (ROP) rotations.
• Existing rotation-invariant models struggle with ROP due to limitations in capturing semantic and topological relationships.
• Current face datasets often have unreliable landmark annotations and lack precise center annotations, hindering model training.

Purpose of the Study:

• To propose a topology-guided semantic face center estimation method robust to RIP and ROP.
• To develop a rotation-aware face dataset with accurate annotations and balanced rotational diversity.
• To introduce a Hybrid-ViT model for accurate landmark localization under extreme poses.

Main Methods:

• Leveraging graph-based landmark relationships for structural integrity preservation.
• Constructing a rotation-aware face dataset with precise center annotations.
• Employing a Hybrid-ViT model fusing CNN and transformer features with a center-guided module.
• Designing a hybrid metric combining topological geometry and semantic perception for evaluation.

Main Results:

• The proposed topology-guided semantic face center estimation method demonstrates superior performance.
• The Hybrid-ViT model achieves robust landmark localization under extreme rotations.
• Experimental results show state-of-the-art performance in cross-dataset evaluations.

Conclusions:

• The developed method effectively addresses challenges in face center estimation under rotational variations.
• The novel dataset and Hybrid-ViT model provide a strong foundation for future research in pose-invariant face analysis.
• The findings offer significant improvements for face detection and recognition systems operating in unconstrained environments.