Linear local models for monocular reconstruction of deformable surfaces

Area of Science:

• Computer Vision
• 3D Geometry
• Machine Learning

Background:

• Recovering 3D shape from a single viewpoint is ambiguous and challenging.
• Existing methods rely on global deformation models learned from data, requiring retraining for each new surface.
• This retraining process is often impractical due to data acquisition needs.

Purpose of the Study:

• To develop a more flexible and practical approach for 3D nonrigid surface shape reconstruction.
• To replace computationally expensive global deformation models with efficient local models.
• To enable 3D shape recovery without the need for retraining for novel surface shapes.

Main Methods:

• Replaced global deformation models with linear local models for surface patches.
• Assembled local models to represent arbitrary surface shapes made of the same material.
• Formulated 3D shape reconstruction as an algebraic problem (closed-form solution) or a convex optimization problem.

Main Results:

• Demonstrated elimination of the need to retrain models for different surface shapes.
• Achieved efficient 3D shape reconstruction using standard numerical packages.
• Presented quantitative results on synthetic data and qualitative results on real images.

Conclusions:

• Linear local models offer a more adaptable and efficient solution for 3D nonrigid surface reconstruction.
• The proposed method overcomes the limitations of global models, reducing the need for extensive data and retraining.
• This approach facilitates practical 3D shape recovery from single-viewpoint data.