3D Face From X: Learning Face Shape From Diverse Sources.

Area of Science:

• Computer Vision
• 3D Computer Graphics
• Machine Learning

Background:

• Traditional 3D face modeling relies on single data sources (scanned data or images).
• 3D scanned data offer geometric accuracy but are costly and limited in subject diversity.
• In-the-wild face images are abundant but lack explicit geometric information.

Purpose of the Study:

• To develop a novel method for jointly learning a 3D face parametric model and 3D face reconstruction.
• To create a unified face model capable of integrating diverse data sources.
• To overcome limitations of single-source 3D face modeling approaches.

Main Methods:

• Proposed a unified learning framework for 3D face modeling.
• Integrated diverse data sources: 3D scanned data, in-the-wild face images, and RGB-D images (iPhone X).
• Developed a method to bridge the geometric information gap between different data modalities.

Main Results:

• Successfully learned a unified 3D face model from diverse data sources.
• Demonstrated improved performance and robustness compared to single-source methods.
• Showcased the effectiveness of incorporating RGB-D data to enhance geometric accuracy.

Conclusions:

• Training 3D face models with diverse data sources significantly enhances their power and accuracy.
• The proposed method effectively unifies information from varied facial data types.
• This approach advances the field of 3D face reconstruction and parametric modeling.