Research map / 3D vision / Face and shape
NeuRodin: A Two-stage Framework for High-Fidelity Neural Surface Reconstruction
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes NeuRodin, a two-stage SDF-based volume rendering framework with local scale SDF-to-density conversion, explicit bias correction, and stochastic-step numerical gradients, achieving higher F-scores on Tanks and Temples and a new ScanNet++ benchmark.
Problem
SDF-based volume rendering for neural surface reconstruction often fails to capture fine geometric detail, producing visible defects due to SDF-to-density conversion bias and over-regularization of geometry.
Approach
Replace the global SDF-to-density scale with a learned local scale, add an explicit loss aligning the maximum-probability distance with the zero level set, use stochastic-step finite-difference gradients in a coarse stage to avoid over-regularization, then refine with Eikonal and smoothness losses and the TUVR density conversion.
Key results
- On the Tanks and Temples training subset, NeuRodin reaches mean F-score 0.51, versus 0.50 for Neuralangelo.
- On the Tanks and Temples advanced subset, mean F-score 28.84 versus 27.14 for the second-best (Bakedangelo, Neuralangelo*).
- On the 8-scene ScanNet++ benchmark, mean F-score 0.638 without priors, above Neuralangelo's 0.564; MonoSDF-Grid with priors scores 0.642.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use NeuRodin instead of Neuralangelo | Tanks and Temples training subset · Tanks and Temples training subset (mean) | F-Score: 0.50 -> 0.51 | Helped |
| Use NeuRodin instead of Neuralangelo* | Tanks and Temples advance subset · Tanks and Temples advance subset (mean) | F-score: 27.14 -> 28.84 | Helped |
| Use NeuRodin instead of MonoSDF-Grid | ScanNet++ · ScanNet++ (mean) | F-score: 0.642 -> 0.638 | Hurt |
| Add local scale for SDF-to-density conversion (ablation) | NeuRodin · Tanks and Temples: Meetingroom | surface accuracy | Helped |
| Add stochastic-step numerical gradient estimation (ablation) | NeuRodin · Tanks and Temples: Meetingroom | surface accuracy | Helped |
| Add explicit bias correction (ablation) | NeuRodin · Tanks and Temples: Meetingroom | surface accuracy | Helped |
| Add stage-two refinement (ablation) | NeuRodin · Tanks and Temples: Meetingroom | surface smoothness | Helped |
| Use estimated normals instead of analytical normals for color conditioning, with stage-one Eikonal loss excluded (ablation) | NeuRodin · Tanks and Temples: Meetingroom | F-score: 0.21 -> 0.35 | Helped |
| Use estimated normals instead of analytical normals for color conditioning, with stage-one Eikonal loss excluded (ablation) | NeuRodin · Tanks and Temples: Courthouse | F-score: 0.11 -> 0.11 | No clear effect |
Surprising findings
- Naively removing geometric regularization or color conditioning on normals leads to unnatural zero level sets, motivating the stochastic-step gradient design.
Limitations
- Explicit bias correction approximates t* by the sampled point with largest weight, introducing deviation.
- Shown to underperform on some scenes (e.g., Courthouse F-score 0.21 vs Neuralangelo 0.28 on the advanced subset is not compared; on training set Courthouse 0.21 is below Neuralangelo's 0.28).
Experimental setup
Datasets: Tanks and Temples ScanNet++
Benchmarks: Tanks and Temples ScanNet++
Metrics: F-score
Baselines: NeuralWarp COLMAP NeuS Geo-NeuS NeuS-NGP MonoSDF Neuralangelo VolSDF
Builds on: NeuS VolSDF Neuralangelo TUVR PermutoSDF
Closest papers by meaning
- Neural Signed Distance Function Inference through Splatting 3D Gaussians Pulled on Zero-Level Set (NeurIPS 2024 · similarity 0.88)
- ProbeSDF: Light Field Probes For Neural Surface Reconstruction (CVPR 2025 · similarity 0.87)
- SC-NeuS: Consistent Neural Surface Reconstruction from Sparse and Noisy Views (AAAI 2024 · similarity 0.86)
- Small Steps and Level Sets: Fitting Neural Surface Models with Point Guidance (CVPR 2024 · similarity 0.86)
- NC-SDF: Enhancing Indoor Scene Reconstruction Using Neural SDFs with View-Dependent Normal Compensation (CVPR 2024 · similarity 0.86)
- PORF: POSE RESIDUAL FIELD FOR ACCURATE NEURAL SURFACE RECONSTRUCTION (ICLR 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.