NeRF-Texture: Synthesizing Neural Radiance Field Textures.

Area of Science:

• Computer Graphics
• Computational Imaging

Background:

• Existing 2D texture synthesis methods struggle with real-world 3D meso-structure textures like grass, leaves, and fabrics.
• Modeling these complex textures requires capturing both geometry and view-dependent appearance.

Purpose of the Study:

• To develop a novel texture synthesis method using Neural Radiance Fields (NeRF) for capturing and generating 3D textures from multi-view images.
• To effectively model and synthesize textures with meso-structure on both planar and curved surfaces.

Main Methods:

• A NeRF-based texture representation disentangles scenes into meso-structure textures and base shapes, learning textures as latent features.
• Latent features are fed into a simultaneously trained NeRF decoder for view-dependent appearance.
• Patch matching of latent features is used for synthesis, regularized by a clustering constraint to enhance quality.

Main Results:

• The proposed method successfully synthesizes NeRF-based textures with fine geometric details and rich appearance.
• It demonstrates effectiveness in handling textures with meso-structure on planar and curved surfaces.
• Evaluations confirm the superior performance of the NeRF-based texture synthesis approach.

Conclusions:

• The novel NeRF texture representation effectively captures and synthesizes complex 3D textures, overcoming limitations of 2D methods.
• The method's ability to handle meso-structure and apply to curved surfaces offers practical advantages for computer graphics.
• Regularization techniques ensure high-quality synthesis by addressing potential inconsistencies in feature spaces.