Research map / 3D vision / NeRF and radiance fields
I2-NeRF: Learning Neural Radiance Fields Under Physically-Grounded Media Interactions
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 I²-NeRF, a physically grounded NeRF for media degradation (underwater, haze, low-light) with reverse-stratified upsampling for near-uniform medium sampling and a Beer–Lambert radiative formulation with vertical downwelling attenuation.
Problem
Standard NeRF assumes clear air and object-centric sampling, which under-samples the participating medium and ignores vertical/horizontal physical attenuation, hurting reconstruction and metric plausibility under media degradation.
Approach
A reverse-stratified upsampling (RSU) strategy that places extra samples in low-density media regions, plus a unified matting-style radiative model (direct + backscatter with Beer–Lambert attenuation) that models downwelling attenuation so the formulation is isotropic across vertical and horizontal directions.
Key results
- On LOM low-light dataset, achieves highest PSNR across all five scenes, e.g. 22.87 on bike vs 22.75 for LITA-GS (second best).
- On SeaThru-NeRF, competitive results: Hazy Fern PSNR 30.59 vs SeaThru-NeRF 30.75; Watersplatting leads several underwater metrics.
- Estimated downwelling water depth of 7.2 m in Okinawa scene vs 7.5 m measured.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use I2-NeRF instead of SeaThru-NeRF | SeaThru-NeRF Hazy Fern · SeaThru-NeRF Hazy Fern | PSNR: 30.75 -> 30.59 | Hurt |
| Use I2-NeRF instead of Watersplatting | SeaThru-NeRF Curaçao · SeaThru-NeRF Curaçao | PSNR: 32.20 -> 32.70 | Helped |
| Use I2-NeRF instead of Watersplatting | SeaThru-NeRF IUI3 Red Sea · SeaThru-NeRF IUI3 Red Sea | PSNR: 29.84 -> 27.33 | Hurt |
| Use I2-NeRF instead of Proposed-T | SeaThru-NeRF J.G. Red Sea · SeaThru-NeRF J.G. Red Sea | PSNR: 25.81 -> 24.11 | Hurt |
| Use I2-NeRF instead of Watersplatting | SeaThru-NeRF Panama · SeaThru-NeRF Panama | PSNR: 31.62 -> 31.55 | Hurt |
| Use I2-NeRF instead of LITA-GS | LOM low-light restoration · LOM sofa | PSNR: 20.43 -> 25.89 | Helped |
| Add media upsampling (ablation) | I2-NeRF on LOM bike · LOM bike | PSNR: 20.92 -> 22.87 | Helped |
| Add geometry loss (ablation) | I2-NeRF on LOM bike · LOM bike | PSNR: 20.66 -> 22.87 | Helped |
| Add compensated similarity loss (ablation) | I2-NeRF on LOM bike · LOM bike | PSNR: 6.55 -> 22.87 | Helped |
| Add mutual exclusion loss (ablation) | I2-NeRF on LOM bike · LOM bike | PSNR: 22.71 -> 22.87 | Helped |
| Add transmittance loss (ablation) | I2-NeRF on LOM bike · LOM bike | PSNR: 21.87 -> 22.87 | Helped |
| Use nonconstant media field (ablation) | I2-NeRF on Okinawa underwater scene · Okinawa underwater scene | PSNR: 31.25 -> 33.97 | Helped |
Surprising findings
- Authors report that ZipNeRF and some baselines collapse to very low PSNR on LOM, while the vanilla ablation without compensated SSIM drops to 6.55 PSNR.
Limitations
- Requires longer training time than rasterization-based methods due to neural implicit representation.
- Assumes static media with fixed optical properties; cannot handle dynamic water or fog.
- Static virtual absorption medium cannot adapt to varying illumination across views in low-light settings.
Experimental setup
Datasets: SeaThru-NeRF LOM
Benchmarks: SeaThru-NeRF LOM
Metrics: PSNR SSIM LPIPS
Baselines: Mip-NeRF 360 Zip-NeRF 3DGS SeaThru-NeRF Proposed-T Watersplatting Aleth-NeRF Luminance-GS LITA-GS SCI+NeRF IAT+NeRF SCI+GS
Builds on: NeRF Zip-NeRF RawNeRF Revised Underwater Image Formation model Bright Channel Prior
Closest papers by meaning
- Decoupling Scattering: Pseudo-Label Guided NeRF for Scenes with Scattering Media (AAAI 2025 · similarity 0.90)
- Aleth-NeRF: Illumination Adaptive NeRF with Concealing Field Assumption (AAAI 2024 · similarity 0.89)
- Rad-NeRF: Ray-decoupled Training of Neural Radiance Field (NeurIPS 2024 · similarity 0.89)
- Physical-aware Neural Radiance Fields for Efficient Exposure Correction (AAAI 2025 · similarity 0.89)
- PBR-NeRF: Inverse Rendering with Physics-Based Neural Fields (CVPR 2025 · similarity 0.88)
- NeRF-VPT: Learning Novel View Representations with Neural Radiance Fields via View Prompt Tuning (AAAI 2024 · 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.