NeuRodin: A Two-stage Framework for High-Fidelity Neural Surface Reconstruction

TL;DR

NeuRodin tackles the challenge of high-fidelity neural surface reconstruction from posed RGB imagery by addressing two key shortcomings of SDF-based volume rendering: SDF-to-density conversion biases and geometry over-regularization. It introduces a local, coordinate-dependent SDF-to-density scale, an explicit bias correction loss that aligns the maximum rendering weight with the SDF zero level set, and a two-stage coarse-to-fine optimization to separate topology changes from smooth surface refinement. The method achieves state-of-the-art reconstruction on indoor and outdoor datasets (Tanks and Temples, ScanNet++) with RGB-only inputs, and it provides a new ScanNet++ benchmark, demonstrating robust detail preservation and topology flexibility. Despite runtime considerations and limitations in textureless regions, NeuRodin offers a practical path toward accurate, topology-flexible 3D surface reconstruction for applications in AR/VR and digital twins.

Abstract

Signed Distance Function (SDF)-based volume rendering has demonstrated significant capabilities in surface reconstruction. Although promising, SDF-based methods often fail to capture detailed geometric structures, resulting in visible defects. By comparing SDF-based volume rendering to density-based volume rendering, we identify two main factors within the SDF-based approach that degrade surface quality: SDF-to-density representation and geometric regularization. These factors introduce challenges that hinder the optimization of the SDF field. To address these issues, we introduce NeuRodin, a novel two-stage neural surface reconstruction framework that not only achieves high-fidelity surface reconstruction but also retains the flexible optimization characteristics of density-based methods. NeuRodin incorporates innovative strategies that facilitate transformation of arbitrary topologies and reduce artifacts associated with density bias. Extensive evaluations on the Tanks and Temples and ScanNet++ datasets demonstrate the superiority of NeuRodin, showing strong reconstruction capabilities for both indoor and outdoor environments using solely posed RGB captures. Project website: https://open3dvlab.github.io/NeuRodin/

Figures (17)

• Figure 1: We present NeuRodin, a novel two-stage framework designed for high-fidelity neural surface reconstruction with intricate structures. Requiring only posed RGB captures as inputs, NeuRodin not only recovers large-scale areas but also accurately reconstructs fine-grained details.
• Figure 2: Comparative analysis of SDF-based and density-based volume rendering methods.(a) Neuralangelo li2023neuralangelo experiences difficulties with topological transformations, leading to incorrect surfaces. (b) Instant-NGP muller2022instant approximates the correct surface positioning yet produces a noisy surface. (c) Our method achieves high-quality surfaces with fine details.
• Figure 3: Visualization of the bias of the density. (a) An ideal scenario where the geometry of the volume rendering scheme (A: maximum probability distance and B: rendered distance) aligns precisely with the geometry of the implicit surface (C: zero level set). (b) A biased scenario showcasing misalignment.
• Figure 4: The heatmap for the variance of the normal predicted using random step.
• Figure 5: Quantitative comparison on the training subset of Tanks and Temples dataset.