ND-SDF: Learning Normal Deflection Fields for High-Fidelity Indoor Reconstruction

Ziyu Tang,Weicai Ye,Yifan Wang,Di Huang,Hujun Bao,Tong He,Guofeng Zhang

2024-09-26

Abstract:Neural implicit reconstruction via volume rendering has demonstrated its effectiveness in recovering dense 3D surfaces. However, it is non-trivial to simultaneously recover meticulous geometry and preserve smoothness across regions with differing characteristics. To address this issue, previous methods typically employ geometric priors, which are often constrained by the performance of the prior models. In this paper, we propose ND-SDF, which learns a Normal Deflection field to represent the angular deviation between the scene normal and the prior normal. Unlike previous methods that uniformly apply geometric priors on all samples, introducing significant bias in accuracy, our proposed normal deflection field dynamically learns and adapts the utilization of samples based on their specific characteristics, thereby improving both the accuracy and effectiveness of the model. Our method not only obtains smooth weakly textured regions such as walls and floors but also preserves the geometric details of complex structures. In addition, we introduce a novel ray sampling strategy based on the deflection angle to facilitate the unbiased rendering process, which significantly improves the quality and accuracy of intricate surfaces, especially on thin structures. Consistent improvements on various challenging datasets demonstrate the superiority of our method.

Computer Vision and Pattern Recognition,Artificial Intelligence

What problem does this paper attempt to address?

The problem that this paper attempts to solve is to perform high - fidelity indoor 3D surface reconstruction under multi - view images. Specifically, existing methods based on neural implicit representation have demonstrated their effectiveness in recovering dense 3D surfaces, but the task of maintaining both geometric details and smoothness remains challenging. Especially in areas with less texture (such as walls and ceilings), relying solely on color images for supervision often leads to under - constrained problems, thus affecting the reconstruction quality. To address this challenge, the paper proposes a new method - ND - SDF (Normal Deflection Signed Distance Function). This method dynamically learns the deviation between the actual scene normal and the prior normal by introducing a "Normal Deflection Field", thereby adaptively using samples in regions with different characteristics and improving the accuracy and effectiveness of the model. ND - SDF can not only obtain smooth weakly - textured regions (such as walls and floors), but also preserve the geometric details of complex structures. In addition, the paper also proposes a new ray - sampling strategy based on the deflection angle to promote an unbiased rendering process, significantly improving the quality and accuracy of complex surfaces, especially the performance on fine structures.

ND-SDF: Learning Normal Deflection Fields for High-Fidelity Indoor Reconstruction

NC-SDF: Enhancing Indoor Scene Reconstruction Using Neural SDFs with View-Dependent Normal Compensation

DebSDF: Delving into the Details and Bias of Neural Indoor Scene Reconstruction

I2-SDF: Intrinsic Indoor Scene Reconstruction and Editing Via Raytracing in Neural SDFs

NeUDF: Leaning Neural Unsigned Distance Fields with Volume Rendering

Error-aware Sampling in Adaptive Shells for Neural Surface Reconstruction

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

Recovering Fine Details for Neural Implicit Surface Reconstruction

Ray-Distance Volume Rendering for Neural Scene Reconstruction

Fine-detailed Neural Indoor Scene Reconstruction using multi-level importance sampling and multi-view consistency

Indoor Scene Reconstruction with Fine-Grained Details Using Hybrid Representation and Normal Prior Enhancement

NeuS: Learning Neural Implicit Surfaces by Volume Rendering for Multi-view Reconstruction

P$^2$SDF for Neural Indoor Scene Reconstruction

NeuralRoom: Geometry-Constrained Neural Implicit Surfaces for Indoor Scene Reconstruction.

Depth-NeuS: Neural Implicit Surfaces Learning for Multi-view Reconstruction Based on Depth Information Optimization

NeuralUDF: Learning Unsigned Distance Fields for Multi-view Reconstruction of Surfaces with Arbitrary Topologies

Edge-aware Neural Implicit Surface Reconstruction

NeuDA: Neural Deformable Anchor for High-Fidelity Implicit Surface Reconstruction

AniSDF: Fused-Granularity Neural Surfaces with Anisotropic Encoding for High-Fidelity 3D Reconstruction

Neural Unsigned Distance Fields for Implicit Function Learning

Learning a Room with the Occ-SDF Hybrid: Signed Distance Function Mingled with Occupancy Aids Scene Representation