Research map / 3D vision / NeRF and radiance fields
GNeRP: Gaussian-guided Neural Reconstruction of Reflective Objects with Noisy Polarization Priors
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 GNeRP, an SDF-based neural surface reconstruction method that models normals as 3D Gaussians splatted to 2D Gaussians supervised by polarization priors, with a DoP-based reweighting to handle noisy polarization cues, and introduces the PolRef dataset.
Problem
Neural surface reconstruction of reflective (specular) objects fails because specular radiance entangles with complex geometry, and polarization priors are noisy in diffuse regions.
Approach
Represent the normal distribution in a neighborhood as a 3D Gaussian over SDF normals, show it splats exactly to a 2D Gaussian in image space, and supervise its mean and covariance with polarization AoP/DoP cues, reweighted per pixel by degree of polarization to balance radiance and polarization losses.
Key results
- GNeRP achieves mean Chamfer distance 1.35 across four PolRef scenes, versus 2.34 for the next best Ref-NeuS and 2.58 for NeuS.
- On Ironman, GNeRP reaches 1.34 vs 1.88 for Ref-NeuS; on Snorlax 1.05 vs 1.99.
- Ablation on Snorlax and Cow shows DoP reweighting improves polarization losses (e.g., L_mean from 2.03 to 1.81 on Snorlax), and full model is best (1.05 and 1.39).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GNeRP's Gaussian normal representation and DoP-reweighted polarization supervision instead of Ref-NeuS | GNeRP · Mean (four PolRef scenes) | Chamfer Distance: 2.34 -> 1.35 | Helped |
| Add covariance supervision and DoP reweighting to naive mean-only supervision (ablation) | GNeRP · Snorlax | Chamfer Distance: 2.03 -> 1.05 | Helped |
| Add covariance supervision and DoP reweighting to naive mean-only supervision (ablation) | GNeRP · Cow | Chamfer Distance: 2.72 -> 1.39 | Helped |
| Add mean supervision and DoP reweighting to covariance-only supervision (ablation) | GNeRP · Snorlax | Chamfer Distance: 3.01 -> 1.05 | Helped |
| Add mean supervision and DoP reweighting to covariance-only supervision (ablation) | GNeRP · Cow | Chamfer Distance: 5.54 -> 1.39 | Helped |
| Use DoP-reweighted mean supervision instead of naive mean supervision (ablation) | GNeRP · Snorlax | Chamfer Distance: 2.03 -> 1.81 | Helped |
| Use DoP-reweighted mean supervision instead of naive mean supervision (ablation) | GNeRP · Cow | Chamfer Distance: 2.72 -> 1.94 | Helped |
| Use DoP-reweighted covariance supervision instead of naive covariance supervision (ablation) | GNeRP · Snorlax | Chamfer Distance: 3.01 -> 2.07 | Helped |
| Use DoP-reweighted covariance supervision instead of naive covariance supervision (ablation) | GNeRP · Cow | Chamfer Distance: 5.54 -> 2.29 | Helped |
Surprising findings
- Naive supervision with covariance-only polarization loss gives the worst scores, indicating detail-focused supervision alone is harmful without the radiance term.
Limitations
- Assumes incident light is unpolarized and objects are directly illuminated.
- Polarization priors remain noisy in diffuse regions; reweighting only partially mitigates this.
- Ground truth available only for four scenes in the dataset; PANDORA lacks ground truth for quantitative evaluation.
Experimental setup
Datasets: PANDORA PolRef (newly captured) Shiny Blender (mentioned, not used)
Benchmarks: PolRef (Ironman, Duck, Cow, Snorlax) PANDORA
Metrics: Chamfer Distance
Baselines: UniSurf VolSDF NeuS Geo-NeuS NeuralWarp NeRO Ref-NeuS PANDORA
Builds on: NeuS NeRF PANDORA Gaussian splatting Mip-NeRF
Closest papers by meaning
- NeRSP: Neural 3D Reconstruction for Reflective Objects with Sparse Polarized Images (CVPR 2024 · similarity 0.91)
- SpecNeRF: Gaussian Directional Encoding for Specular Reflections (CVPR 2024 · similarity 0.89)
- UniSDF: Unifying Neural Representations for High-Fidelity 3D Reconstruction of Complex Scenes with Reflections (NeurIPS 2024 · similarity 0.89)
- Normal-NeRF: Ambiguity-Robust Normal Estimation for Highly Reflective Scenes (AAAI 2025 · similarity 0.89)
- NeRFPrior: Learning Neural Radiance Field as a Prior for Indoor Scene Reconstruction (CVPR 2025 · similarity 0.88)
- HyRF: Hybrid Radiance Fields for Memory-efficient and High-quality Novel View Synthesis (NeurIPS 2025 · similarity 0.88)
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.