Research map / 3D vision / NeRF and radiance fields
Normal-NeRF: Ambiguity-Robust Normal Estimation for Highly Reflective Scenes
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 a transmittance-gradient normal estimator and dual activated densities (softplus and exp) for NeRF so that reflective scenes get robust normals and sharp boundaries, improving reflection-aware rendering on synthetic and real datasets.
Problem
NeRF-based normal estimation is unreliable on highly reflective scenes because density gradients become irregular under shape ambiguity (semi-transparent fake reflections), and existing fixes rely on geometry priors that oversmooth complex geometry.
Approach
Estimate normals from the gradient of the monotonic accumulated transmittance rather than the density gradient, and use two activations on a shared density MLP (exp for sharp rendering density, softplus for smooth density used in normal computation), plus a stop-gradient warmup on the predicted-normal loss.
Key results
- On Shiny Blender the method reaches 39.24 PSNR, 0.982 SSIM and 0.040 LPIPS, beating the best baseline Ref-NeRF (35.96 PSNR).
- On Glossy Synthetic it achieves 33.24 PSNR and 0.971 SSIM, versus 30.49 PSNR for the second-best Ref-NeuS.
- On NeRF Synthetic it reports MAE 14.486 versus 20.549 for ENVIDR, while matching Zip-NeRF on PSNR (34.53 vs 33.69) and being on par on real captured scenes (24.09 average PSNR).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use transmittance gradient instead of density gradient for normal estimation (ablation) | Normal-NeRF · Glossy Synthetic | PSNR: 27.82 -> 30.67 | Helped |
| Add stop-gradient warmup to the normal loss (ablation) | Normal-NeRF · Glossy Synthetic | PSNR: 30.67 -> 33.24 | Helped |
| Use dual softplus and exp density activations instead of a single exp activation (ablation) | Normal-NeRF · Glossy Synthetic | PSNR: 32.20 -> 33.24 | Helped |
| Use dual softplus and exp density activations instead of a single softplus activation (ablation) | Normal-NeRF · Glossy Synthetic | PSNR: 32.79 -> 33.24 | Helped |
| Use exponential lambda_n warmup instead of constant lambda_n (ablation) | Normal-NeRF · Shiny Blender | PSNR: 36.40, 36.82, 36.78, 36.35 -> 37.02 | Helped |
| Use reflected direction instead of viewing direction in the environment MLP (ablation) | Normal-NeRF · Glossy Synthetic | PSNR: 25.41 -> 33.24 | Helped |
| Use predicted normals for reflection directions rather than transmittance gradients directly (ablation) | Normal-NeRF · Glossy Synthetic | PSNR: 25.90 -> 33.24 | Helped |
| Add frequency positional encoding for normal prediction (ablation) | Normal-NeRF · Glossy Synthetic | PSNR: 32.58 -> 33.24 | Helped |
| Add hash-grid features for normal prediction (ablation) | Normal-NeRF · Glossy Synthetic | PSNR: 30.41 -> 33.24 | Helped |
| Apply normalized weight decay to the hash grid (ablation) | Normal-NeRF · Glossy Synthetic | PSNR: 30.26 -> 33.24 | Helped |
| Add transmittance-gradient normal estimation (ablation) | Normal-NeRF · Glossy Synthetic | PSNR: 27.82 -> 33.24 | Helped |
| Add stop-gradient warmup (ablation) | Normal-NeRF · Glossy Synthetic | PSNR: 31.64 -> 33.24 | Helped |
Surprising findings
- Ablation shows that using only a softplus density hurts thin structure reconstruction, and a single exp density undermines normal estimation robustness.
- The predicted normal loss can hinder density optimization, so the stop-gradient warmup is needed; constant lambda_n settings all underperform the warmup schedule.
Limitations
- On real captured scenes the method performs only on par with existing methods, with Zip-NeRF better on LPIPS.
- The transmittance gradient is less precise, so a spatial MLP is needed to refine normals, adding a learned component.
- Ref-NeRF and NMF numbers on some datasets are taken from their papers rather than rerun.
Experimental setup
Datasets: NeRF Synthetic Shiny Blender Glossy Synthetic Ref-NeRF real captured scenes
Benchmarks: NeRF Synthetic Shiny Blender Glossy Synthetic Ref-NeRF real captured scenes
Metrics: PSNR SSIM LPIPS MAE
Baselines: Zip-NeRF Ref-NeRF NMF Ref-NeuS ENVIDR
Builds on: NeRF Ref-NeRF Instant-NGP Zip-NeRF Nerfstudio
Closest papers by meaning
- GNeRP: Gaussian-guided Neural Reconstruction of Reflective Objects with Noisy Polarization Priors (ICLR 2024 · similarity 0.89)
- NeRFPrior: Learning Neural Radiance Field as a Prior for Indoor Scene Reconstruction (CVPR 2025 · similarity 0.88)
- UniSDF: Unifying Neural Representations for High-Fidelity 3D Reconstruction of Complex Scenes with Reflections (NeurIPS 2024 · similarity 0.87)
- NeRSP: Neural 3D Reconstruction for Reflective Objects with Sparse Polarized Images (CVPR 2024 · similarity 0.87)
- NeRF On-the-go: Exploiting Uncertainty for Distractor-free NeRFs in the Wild (CVPR 2024 · similarity 0.87)
- PBR-NeRF: Inverse Rendering with Physics-Based Neural Fields (CVPR 2025 · similarity 0.87)
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.