Pose-Invariant Face Recognition via RGB-D Images

Area of Science:

• Computer Vision
• Biometrics
• Pattern Recognition

Background:

• Two-dimensional (2D) face recognition struggles with large pose variations, self-occlusion, and deformation.
• Three-dimensional (3D) face models offer inherent advantages for handling pose variations.
• RGB-D imaging provides both texture and depth information, enabling richer facial representations.

Purpose of the Study:

• To propose a novel pose-invariant face recognition method utilizing RGB-D images.
• To leverage depth information to overcome limitations of 2D face recognition, specifically self-occlusion and deformation.
• To enhance face recognition performance under challenging conditions with significant pose variations.

Main Methods:

• Utilizing depth data from RGB-D images to render gallery images to match probe views.
• Employing depth information for similarity measurement through techniques like frontalization and symmetric filling.
• Integrating both texture and depth information for robust identity estimation.

Main Results:

• Demonstrated improved performance in face recognition across multiple benchmark datasets (Bosphorus, CurtinFaces, Eurecom, Kiwi).
• Showcased the effectiveness of depth information in handling large pose variations and challenging environmental conditions.
• Validated the method's ability to mitigate issues like self-occlusion and deformation inherent in 2D recognition.

Conclusions:

• The proposed RGB-D based method significantly enhances pose-invariant face recognition.
• Depth information is crucial for robust face recognition, especially under large pose variations and occlusions.
• This approach offers a promising direction for developing more accurate and resilient biometric systems.