Scalable and High-Quality Neural Implicit Representation for 3D Reconstruction.

Area of Science:

• Computer Vision
• Computer Graphics
• Machine Learning

Background:

• Neural implicit representations, particularly signed distance functions (SDFs), show promise for 3D surface reconstruction.
• Existing global SDF-based methods face limitations in accuracy and scalability due to single network constraints.

Purpose of the Study:

• To develop a versatile, scalable, and high-quality neural implicit representation for 3D surface reconstruction.
• To overcome the limitations of existing global methods in terms of accuracy and scale.

Main Methods:

• Proposed a divide-and-conquer approach by modeling objects/scenes as a fusion of multiple local neural SDFs with overlapping regions.
• Key steps include: (1) determining local field distribution and overlap, (2) relative pose registration of local SDFs, and (3) SDF blending.
• Leveraged object structure or data distribution for local field construction.

Main Results:

• Achieved high-fidelity surface reconstruction through independent local region representation.
• Enabled scalable scene reconstruction, addressing limitations of previous global methods.
• Experimental results validated the method's effectiveness and practicality.

Conclusions:

• The proposed fusion of local neural SDFs offers a robust solution for accurate and scalable 3D reconstruction.
• This approach enhances the modeling capabilities of neural implicit surfaces.
• Demonstrated practical applicability in complex 3D scene reconstruction tasks.