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

Ziyi Chen,Xiaolong Wu,Yu Zhang

2024-05-01

Abstract:State-of-the-art neural implicit surface representations have achieved impressive results in indoor scene reconstruction by incorporating monocular geometric priors as additional supervision. However, we have observed that multi-view inconsistency between such priors poses a challenge for high-quality reconstructions. In response, we present NC-SDF, a neural signed distance field (SDF) 3D reconstruction framework with view-dependent normal compensation (NC). Specifically, we integrate view-dependent biases in monocular normal priors into the neural implicit representation of the scene. By adaptively learning and correcting the biases, our NC-SDF effectively mitigates the adverse impact of inconsistent supervision, enhancing both the global consistency and local details in the reconstructions. To further refine the details, we introduce an informative pixel sampling strategy to pay more attention to intricate geometry with higher information content. Additionally, we design a hybrid geometry modeling approach to improve the neural implicit representation. Experiments on synthetic and real-world datasets demonstrate that NC-SDF outperforms existing approaches in terms of reconstruction quality.

Computer Vision and Pattern Recognition

What problem does this paper attempt to address?

The problem that this paper attempts to solve is the degradation of reconstruction quality in high - quality indoor scene reconstruction in multi - view images due to the multi - view inconsistency among monocular geometric priors. Specifically, although the existing neural implicit surface representation methods have achieved remarkable results in indoor scene reconstruction, when dealing with indoor scenes containing a large number of texture - less regions, the performance of these methods will be affected because the multi - view photometric consistency cannot provide sufficient constraints. In addition, there is a deviation between the geometric priors predicted by the monocular geometric estimation network and the true values, and this deviation is related to the viewing angle, which further affects the consistency and details of the reconstruction.

To overcome these problems, the paper proposes NC - SDF (Neural SDF with View - Dependent Normal Compensation), which is a 3D reconstruction framework of neural signed distance field (SDF) combined with view - dependent normal compensation. By adaptively learning and correcting the deviation in the monocular normal priors, NC - SDF can reduce the negative impact of inconsistent supervision, thereby improving the overall consistency and local details of the reconstruction. At the same time, the paper also introduces an information - pixel sampling strategy to pay more attention to complex geometric structures with high information content, and designs a hybrid geometric modeling method to improve the neural implicit representation. Experimental results show that the reconstruction quality of NC - SDF on both synthetic datasets and real - world datasets is better than that of existing methods.