DeepSketchHair: Deep Sketch-Based 3D Hair Modeling

Area of Science:

• Computer Graphics
• Artificial Intelligence
• 3D Modeling

Background:

• Generating realistic 3D hair models is complex.
• Existing methods often require specialized skills or extensive manual input.
• Bridging the gap between 2D sketching and 3D asset creation is a significant challenge.

Purpose of the Study:

• To introduce DeepSketchHair, a deep learning system for automatic 3D hair modeling from 2D sketches.
• To enable users to create and edit 3D hairstyles intuitively using simple sketches.
• To provide a method for generating 3D hair that accurately reflects user-defined contours and growth directions.

Main Methods:

• Developed a system utilizing three neural networks: S2ONet, O2VNet, and V2VNet.
• S2ONet converts 2D sketches into a dense 2D hair orientation field.
• O2VNet maps the 2D field to a 3D vector field, and V2VNet refines it for editing.
• Trained networks using synthetic data from a 3D hairstyle database.

Main Results:

• Successfully generated 3D hair models that accurately match input 2D sketches.
• Demonstrated the tool's effectiveness and expressiveness across various hairstyles.
• Showcased the capability for iterative hair editing through new sketches from different views.

Conclusions:

• DeepSketchHair offers an effective and intuitive approach to 3D hair modeling.
• The system leverages deep learning to translate simple 2D input into complex 3D assets.
• This method advances the field of character creation and digital asset generation.