LARNeXt: End-to-End Lie Algebra Residual Network for Face Recognition

Area of Science:

• Computer Vision
• Machine Learning
• Biometrics

Background:

• Face recognition is challenged by significant pose variations, especially between frontal and profile views.
• Traditional methods rely on data synthesis or pose-invariant learning, with limitations in handling extreme variations.

Purpose of the Study:

• To develop a pose robust face recognition system using an integrated end-to-end architecture.
• To theoretically and experimentally validate a novel approach to address pose variations in deep feature generation.

Main Methods:

• Proposed a Lie algebra residual architecture (LARNeXt) for end-to-end face recognition.
• Demonstrated that 3D face rotation is equivalent to an additive residual component in CNN feature space.
• Designed three subnets: soft regression for pose estimation, residual for rotation decoding, and gating for residual component control.

Main Results:

• Quantitative and visualization experiments confirmed the theoretical findings and network design effectiveness.
• The LARNeXt method consistently outperformed state-of-the-art methods on various face recognition datasets.
• Evaluations included frontal-profile, unconstrained, and industrial-grade face recognition tasks.

Conclusions:

• The proposed LARNeXt architecture offers a robust solution for pose-variant face recognition.
• The theoretical framework linking face rotation to feature space residuals provides a new perspective.
• The method shows significant potential for real-world applications requiring reliable face recognition across diverse poses.