Shape Synthesis from Sketches via Procedural Models and Convolutional Networks.

Area of Science:

• Computer Graphics
• Computational Design
• Artificial Intelligence

Background:

• Procedural modeling offers high-quality visual content generation via complex rule sets.
• User control over procedural modeling is challenging due to numerous non-linear parameters.

Purpose of the Study:

• To develop an intuitive sketch-based interface for procedural modeling.
• To enable users to easily control procedural model outputs for design purposes.

Main Methods:

• A novel algorithm translates user-provided 2D sketches into procedural model parameters.
• A deep Convolutional Neural Network (CNN) is trained to map sketches to parameters using synthetic data.
• The system generates multiple detailed shapes resembling the input sketch.

Main Results:

• The sketch-based approach effectively bypasses manual parameter tuning.
• Users can explore design variations by modifying sketches or selecting generated outputs.
• The method demonstrates accuracy and efficacy across diverse procedural modeling scenarios.

Conclusions:

• Sketch-based input offers an intuitive alternative to complex parameter manipulation in procedural modeling.
• This approach empowers users to create high-quality procedural content with greater ease.
• The technique is applicable to both man-made and organic shape design.