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

TL;DR

Indoor scene reconstruction with neural implicit surfaces is challenged by multi-view inconsistency in monocular priors. NC-SDF introduces a view-dependent normal compensation model that learns and corrects biases in monocular normals, integrating them into the neural SDF ($SDF$) and radiance fields and producing compensated normals $N^{comp}$. It also adds an informative pixel sampling strategy and a hybrid geometry model to capture fine geometry without sacrificing smoothness. Across real (ScanNet) and synthetic (ICL-NUIM) datasets, NC-SDF achieves state-of-the-art results, demonstrating improved global consistency and sharper details under noisy priors, with practical implications for AR/VR and robotics.

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.

Figures (10)

• Figure 1: Comparison between baseline and NC-SDF. State-of-the-art neural implicit surface representations produce suboptimal reconstructions with noisy or missing surfaces, primarily due to multi-view inconsistency between monocular geometric priors. Our NC-SDF introduces a view-dependent normal compensation model to adaptively learn and correct biases in normal priors. This approach enables the recovery of intricate geometric details while ensuring smoothness in texture-less areas within reconstructions.
• Figure 2: Pipeline of NC-SDF. We model the geometry field (SDF), view-dependent radiance field, and view-dependent normal biases with neural implicit functions. Besides, we propose an informative pixel sampling strategy and a hybrid geometry model to further improve the reconstruction of thin geometry.
• Figure 3: Visualization of rendered results, comparing (a) without and (b) with our normal compensation model.
• Figure 4: Visualization of texture intensity maps at different intensity thresholds. The Canny operator exhibits a more robust performance than the Sobel operator.
• Figure 5: Qualitative results on ScanNet dai2017scannet.