Research map / 3D vision / NeRF and radiance fields
Decoupling Scattering: Pseudo-Label Guided NeRF for Scenes with Scattering Media
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 DecoNeRF, a NeRF with separate object and scattering-medium densities that uses early-training pseudo-labels from a dehazing network plus a cyclical progressive dimensional optimization strategy (CPDOS) to decouple haze/water from objects.
Problem
NeRF reconstruction and novel view synthesis in scenes with scattering media (haze, fog, underwater) is under-constrained, so object and medium densities are hard to decouple and models often collapse to trivial solutions.
Approach
Use pseudo-ground-truth clear images from an existing dehazing network (RQ-LLIE codebook model) as MSE supervision only during the early training stage, combined with a cyclical scheme that alternately optimizes the clear image, transmission map, and atmospheric light (coordinate-descent inspired) to narrow the solution space.
Key results
- On synthetic hazy LLFF scenes, the method reaches 21.42 PSNR, 0.740 SSIM, 0.166 LPIPS, outperforming the best baseline MITNet+NeRF (19.92 PSNR) and SeaThru-NeRF (15.88 PSNR).
- Ablation: removing CPDOS drops PSNR to 19.49; removing label guidance gives 19.55; the full model reaches 21.42 and reduces transmission MSE from 0.0118 to 0.0036.
- Pseudo-label guidance lifts DehazeFormer pseudo-labels from 19.11 to 21.74 PSNR and Codebook pseudo-labels from 18.72 to 22.23 PSNR.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add CPDOS (ablation) | NeRF for hazy scenes · Hazy scenes (LLFF) | PSNR: 19.49 -> 21.42 | Helped |
| add CPDOS (ablation) | NeRF for hazy scenes · Hazy scenes (LLFF) | MSE transmission map: 0.0118 -> 0.0036 | Helped |
| add CPDOS (ablation) | NeRF for hazy scenes · Hazy scenes (LLFF) | MSE atmospheric light: 0.0069 -> 0.0011 | Helped |
| add early pseudo-label guidance (ablation) | NeRF for hazy scenes · Hazy scenes (LLFF) | PSNR: 19.55 -> 21.42 | Helped |
| add early pseudo-label guidance (ablation) | NeRF for hazy scenes · Hazy scenes (LLFF) | MSE transmission map: 0.0132 -> 0.0036 | Helped |
| add early pseudo-label guidance (ablation) | NeRF for hazy scenes · Hazy scenes (LLFF) | MSE atmospheric light: 0.0067 -> 0.0011 | Helped |
| use DehazeFormer-generated pseudo-labels (ablation) | Ours · Hazy scenes (LLFF) | PSNR: 19.11 -> 21.74 | Helped |
| use Codebook-generated pseudo-labels (ablation) | Ours · Hazy scenes (LLFF) | PSNR: 18.72 -> 22.23 | Helped |
Surprising findings
- Using the ground-truth clear image as a regularizer tends to converge to a local optimum, whereas CPDOS escapes it.
- The method surpasses the quality of its own pseudo-labels, removing residual haze that the pseudo-labels retained.
Limitations
- Performance depends on the accuracy of the pseudo-labels, and there is no automated way to assess their adequacy.
- Evaluation covers only a limited set of scenes; real underwater results are qualitative only.
- Hazy data is synthetic (ASM model on LLFF) rather than captured multi-view hazy scenes.
Experimental setup
Datasets: LLFF SeaThru-NeRF underwater data
Benchmarks: LLFF synthetic haze scenes SeaThru-NeRF real underwater scenes
Metrics: PSNR SSIM LPIPS MSE of transmission map MSE of atmospheric light
Baselines: SeaThru-NeRF DCP+NeRF ROP+NeRF DehazeFormer+NeRF MITNet+NeRF Codebook+NeRF Mip-NeRF 360
Builds on: Mip-NeRF 360 SeaThru-NeRF RQ-LLIE Atmospheric scattering model
Closest papers by meaning
- I2-NeRF: Learning Neural Radiance Fields Under Physically-Grounded Media Interactions (NeurIPS 2025 · similarity 0.90)
- Neural Underwater Scene Representation (CVPR 2024 · similarity 0.89)
- Rad-NeRF: Ray-decoupled Training of Neural Radiance Field (NeurIPS 2024 · similarity 0.89)
- PBR-NeRF: Inverse Rendering with Physics-Based Neural Fields (CVPR 2025 · similarity 0.88)
- ReconFusion: 3D Reconstruction with Diffusion Priors (CVPR 2024 · similarity 0.88)
- Optimize the Unseen - Fast NeRF Cleanup with Free Space Prior (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.