Inequality-Constrained and Robust 3D Face Model Fitting.

Area of Science:

• Computer Vision
• 3D Reconstruction
• Geometric Modeling

Background:

• 3D Morphable Models (3DMMs) are crucial for facial shape representation.
• Existing unconstrained optimization methods for 3DMM fitting struggle with parameter tuning and universal weight determination.
• A need exists for robust and versatile 3DMM fitting techniques applicable in diverse conditions.

Purpose of the Study:

• To propose a new formulation for 3DMM fitting using inequality-constrained optimization.
• To eliminate the requirement for weight parameter tuning in 3DMM fitting.
• To enhance the accuracy and robustness of 3DMM fitting, especially for in-the-wild scenarios.

Main Methods:

• Formulated 3DMM fitting as an inequality-constrained optimization problem.
• Incorporated learned intervals for basis coefficients as primary constraints.
• Utilized sparse landmark detectors and a robust Gradient Correlation objective function for in-the-wild operation.

Main Results:

• Achieved performance comparable to deep learning methods on in-the-wild data.
• Outperformed deep learning methods on controlled data with exact ground truth.
• Demonstrated versatility across multiple frames and arbitrary sizes due to its learning-free nature.

Conclusions:

• The proposed inequality-constrained optimization offers a versatile and parameter-free approach to 3DMM fitting.
• This method shows promise for improving facial shape accuracy in various conditions.
• Encourages further exploration of inequality-constrained optimization for 3DMM fitting.