Robust 3D Face Reconstruction Using One/Two Facial Images

Area of Science:

• Computer Vision
• Biometrics
• Machine Learning

Background:

• 3D face reconstruction from 2D images is crucial for face analysis and recognition applications.
• 3D facial data offers advantages over 2D data by being less sensitive to lighting and pose variations.
• Convolutional Neural Networks (CNNs) have advanced 2D to 3D facial reconstruction capabilities.

Purpose of the Study:

• To propose a novel CNN-based method for robust 3D facial reconstruction using two input images (front and side).
• To enhance 3D face reconstruction capabilities for scenarios where only a single 2D image is available.
• To improve the accuracy of 3D face reconstruction compared to current state-of-the-art methods.

Main Methods:

• Developed a novel CNN architecture for 3D face reconstruction from a pair of facial images.
• Trained the CNN model using a combination of synthetic and real facial datasets.
• Introduced an auxiliary network to generate a rotated image for single-image reconstruction scenarios.

Main Results:

• The proposed CNN method achieved higher accuracy in 3D face prediction on the MICC Florence dataset than existing state-of-the-art approaches.
• Demonstrated the effectiveness of the single-image reconstruction scheme, leveraging the auxiliary network.
• Validated the robustness of the 3D reconstruction technique across different facial data types.

Conclusions:

• The novel CNN-based approach significantly advances 3D face reconstruction accuracy from 2D images.
• The method provides a practical solution for law enforcement agencies (LEAs) and other applications requiring 3D face data.
• The single-image reconstruction capability broadens the applicability of the proposed technique in real-world scenarios.