Three-Dimensional Reconstruction of Indoor Scenes Based on Implicit Neural Representation.

Area of Science:

• Computer Vision
• Computer Graphics
• 3D Scene Reconstruction

Background:

• Traditional 3D reconstruction methods struggle with surface details, large textures, and uneven lighting in indoor scenes.
• Existing methods often produce noisy 3D models with floating debris.

Purpose of the Study:

• To develop an advanced 3D reconstruction method for indoor scenes.
• To overcome limitations of traditional methods in detail preservation and noise reduction.

Main Methods:

• Combines neural radiance fields (NeRFs) with signed distance function (SDF) implicit representations.
• Utilizes NeRF volume density for SDF geometric information.
• Incorporates adaptive normal prior optimization for enhanced shape and surface learning.
• Introduces a novel regularization term for weight distribution and noise removal.

Main Results:

• Achieves high-quality geometric information preservation from NeRFs.
• Generates explicit meshes with smooth surfaces using SDFs.
• Significantly improves reconstruction of large plane textures and uneven illumination areas.
• Effectively removes floating debris noise from reconstructed models.

Conclusions:

• The proposed NeRF-SDF hybrid method offers superior 3D indoor scene reconstruction.
• Demonstrates state-of-the-art performance on ScanNet, Hypersim, and Replica datasets.
• Presents a robust solution for detailed and clean 3D indoor scene modeling.