68 landmarks are efficient for 3D face alignment: what about more?: 3D face alignment method applied to face...

Area of Science:

• Computer Vision and Pattern Recognition
• Biometrics and Human-Computer Interaction

Background:

• Accurate 3D face reconstruction from 2D images is crucial for robust face recognition, especially with unconstrained "in the wild" data.
• Existing methods often struggle with noisy landmarks and require 3D Morphable Models, limiting their applicability.

Purpose of the Study:

• To propose a novel 3D face alignment technique for single 2D profile images with noisy landmarks.
• To achieve accurate individual recognition using deep learning on reconstructed 3D face models.

Main Methods:

• Extraction of over 68 facial landmarks using a bag-of-features approach to capture visible and invisible keypoints.
• Reconstruction of a 3D face model and triangular mesh from keypoints, employing butterfly and BPA algorithms for regional correlation.
• Alignment and pose correction by fitting the rendered 3D model to the 2D face, followed by pose normalization for recognition using Deep Convolutional Neural Networks (DCNNs).

Main Results:

• Successful reconstruction of high-quality 3D face models from 2D images without external 3D Morphable Models, leveraging 2D-to-3D annotations.
• Achieved comparable or superior face recognition performance on standard benchmarks (YTF, LFW, BIWI) compared to state-of-the-art methods.
• Demonstrated robustness to noisy landmarks in "in the wild" face images.

Conclusions:

• The proposed 2D-to-3D face alignment and reconstruction method effectively handles noisy landmarks for improved 3D face modeling.
• The integration with DCNNs enables highly accurate face recognition from single profile images, outperforming existing approaches on tested datasets.