Dense 3D Face Correspondence.

Area of Science:

• Computer Vision
• 3D Geometry Processing
• Biometrics

Background:

• Establishing dense correspondences between 3D faces is crucial for tasks like facial recognition and analysis.
• Existing methods often struggle with automation, accuracy, and generalization across diverse datasets.

Purpose of the Study:

• To develop an automated algorithm for establishing dense 3D face correspondences.
• To construct a deformable model from these correspondences for fitting unseen faces.
• To evaluate the accuracy and generalization capabilities of the proposed methods.

Main Methods:

• Automatic detection of sparse correspondences on 3D face boundaries.
• Iterative expansion of correspondences using surface curvature and geodesic curves.
• Construction of a deformable model (K3DM) from dense corresponded faces.
• Iterative fitting of the deformable model to unseen faces via rigid alignment and morphing.

Main Results:

• Achieved dense correspondences with a mean localization error of 1.28 mm on synthetic data.
• Detected 14 anthropometric landmarks on real faces with 3 mm precision.
• Deformable model achieved 98.5% face recognition accuracy on FRGCv2 and 98.6% on Bosphorus.
• Demonstrated generalization capabilities to unseen datasets.

Conclusions:

• The proposed algorithm effectively establishes dense 3D face correspondences automatically.
• The deformable model fitting approach is robust and achieves high accuracy in face recognition.
• The method shows strong potential for various 3D facial analysis applications.