Detail-preserving controllable deformation from sparse examples.

Area of Science:

• Computer Graphics
• Geometric Modeling
• Machine Learning

Background:

• Laser scanning captures intricate static details of real-world objects.
• Faithful digital models require reproducing both static details and dynamic deformations.

Purpose of the Study:

• Develop a data-driven model for synthesizing realistic, deformable digital geometries.
• Capture and reproduce high-resolution details during large-scale deformations.

Main Methods:

• A two-component data-driven model: one for large-scale deformations, another for high-resolution details.
• Nonlinear mappings between sparse control points and bone transformations for global deformations.
• Local mappings in geometry and pose space for realistic synthesis from sparse data.
• Second nonlinear mapping for per-vertex displacements to generate deformable fine-scale details.

Main Results:

• The model effectively synthesizes highly deformable models with rich fine features.
• Demonstrated robustness and effectiveness on scanned human hands, faces, and dinosaur models.
• Successful application in both keyframe and performance-driven animation.

Conclusions:

• The developed data-driven approach enables realistic synthesis of deformable models with intricate details.
• Learned models from sparse data are effective and robust for various applications.
• The method offers an improvement over alternative techniques for deformable model synthesis.