Stylizing Sparse-View 3D Scenes With Hierarchical Neural Representation.

Area of Science:

• Computer Vision
• Computer Graphics
• Artificial Intelligence

Background:

• 3D scene stylization aims to render consistent stylized views from novel angles.
• Neural Radiance Fields (NeRF) are commonly used for 3D scene reconstruction and stylization.
• Few-shot NeRFs struggle with high-frequency artifacts when using sparse input views.

Purpose of the Study:

• To develop a method for high-quality 3D scene stylization directly from sparse input views.
• To address the limitations of existing NeRF-based stylization methods that produce artifacts with limited data.
• To disentangle scene content semantics from style textures for improved stylization.

Main Methods:

• Proposes a coarse-to-fine sparse-view 3D scene stylization framework.
• Introduces a novel hierarchical encoding-based neural representation for direct optimization.
• Implements a new optimization strategy with content strength annealing for realistic stylization and content preservation.

Main Results:

• The proposed method generates high-quality stylized scenes from sparse input views.
• Achieves superior stylization quality compared to fine-tuning-based baselines.
• Demonstrates improved efficiency in sparse-view scene stylization.

Conclusions:

• Directly optimizing encoding-based scene representations is effective for sparse-view stylization.
• The proposed framework successfully disentangles content and style, preserving scene fidelity.
• This approach offers a more robust and efficient solution for 3D scene stylization with limited input data.