Caricature Expression Extrapolation Based on Kendall Shape Space Theory.

Area of Science:

• Computer Vision
• Computer Graphics
• Geometric Deep Learning

Background:

• Facial expression editing is crucial for digital media, but extrapolation (exaggerating beyond original expressions) is underexplored.
• Current 2-D caricature editing primarily uses interpolation, limiting creative possibilities.

Purpose of the Study:

• To propose a novel method for 2-D caricature facial expression extrapolation.
• To introduce a 3-D model representation within the Kendall shape space to handle rigid transformations.

Main Methods:

• Utilizing the Kendall shape space with a novel 3-D model representation to remove translation, scaling, and rotation.
• Reconstructing a 3-D model from a 2-D caricature and extrapolating expressions using exponential mapping in Riemannian space.
• Generating exaggerated caricature expressions based on the extrapolated 3-D facial model.

Main Results:

• The method effectively and automatically extrapolates facial expressions in caricatures with high consistency and fidelity.
• Generated 3-D facial models with diverse expressions, expanding the FaceWarehouse database.
• Demonstrated robustness to facial poses and avoidance of complex 3-D caricature training sets, unlike deep learning methods.

Conclusions:

• The proposed Kendall shape space method offers a robust and efficient approach for facial expression extrapolation in 2D caricatures.
• This technique provides a valuable tool for digital content creation, particularly in film and gaming.
• The method's reliance on standard datasets and avoidance of complex training data makes it a practical alternative to deep learning approaches.