PGT-NeuS: Progressive-Growing Tri-Plane Representation for Neural Surface Reconstruction

Area of Science:

• Computer Graphics
• Computer Vision
• Geometric Modeling

Background:

• Neural 3D reconstruction methods like NeuS offer improvements over traditional techniques.
• Current neural methods struggle with fine geometric detail due to limitations in expressing high-frequency signals.

Purpose of the Study:

• To enhance the reconstruction of fine-grained geometric details in 3D models from multi-view images.
• To improve the stability and reduce the difficulty of neural 3D reconstruction training.
• To address challenges posed by sparse viewpoints and inconsistent lighting in 3D reconstruction.

Main Methods:

• A multi-resolution tri-plane feature encoding is proposed, combining high-resolution detail with low-resolution smoothness.
• A progressive training strategy is implemented, gradually incorporating scene details from coarse to fine.
• Normal priors are introduced as supervision, with consistency verification for multi-view normal priors.
• A perturbing and fine-tuning strategy is applied to unreliable normal prior regions.

Main Results:

• The proposed multi-resolution encoding effectively reconstructs fine geometric details while suppressing artifacts.
• The progressive training strategy enhances reconstruction quality and maintains training stability.
• Normal prior supervision and consistency verification improve surface reconstruction accuracy, especially with sparse data.
• The overall approach demonstrates superior performance in detailed 3D reconstruction.

Conclusions:

• The novel encoding and training strategy significantly advance neural 3D reconstruction capabilities.
• The method offers a robust solution for detailed geometric reconstruction, even in challenging imaging conditions.
• This work provides a foundation for more accurate and stable neural 3D reconstruction systems.