Research map / 3D vision / NeRF and radiance fields
Neural Underwater Scene Representation
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.
Extends NeRF to underwater scenes by modeling wavelength-dependent water attenuation, a log-space illumination field with a shared tone mapper, and a separate dynamic branch for moving objects.
Problem
NeRF assumes static scenes in air, but underwater scenes have distance-dependent light attenuation, unstable local illumination, and moving marine objects, which existing underwater NeRFs (e.g., SeaThru-NeRF) do not jointly handle.
Approach
Treats water as a semi-transparent wavelength-dependent medium inside volume rendering, adds a log-space illumination field with a shared tone mapper for exposure changes, and reconstructs static and dynamic branches in two stages with a mixed progressive sampling strategy.
Key results
- On the SeaThru dataset, Proposed-T reaches 30.03 PSNR on CURACAO and 25.81 on JAPANESE, outperforming SeaThru-NeRF on most scenes.
- On the proposed dataset, the full method reaches 26.17 PSNR on CORAL versus 23.89 for SeaThru-NeRF.
- In a user study with 122 users, the method scored 4.46 realism, 4.77 fidelity, and 4.29 multi-view consistency, the highest of the compared methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the proposed method instead of Instant-NGP, SeaThru-NeRF, or DynamicNeRF | underwater scene reconstruction · CORAL, TURTLE, COMPOSITE, and SARDINE | PSNR, SSIM, and LPIPS | Helped |
| Remove time-dependent network components (ablation) | Proposed · COMPOSITE | visual fidelity | Hurt |
| Remove the illumination field and tone mapper (ablation) | Proposed · COMPOSITE | visual fidelity | Hurt |
| Remove the illumination field but retain the tone mapper (ablation) | Proposed · COMPOSITE | visual fidelity | Hurt |
| Use only the dynamic branch, removing the static branch (ablation) | Proposed · COMPOSITE | visual fidelity | Hurt |
| Remove the tone mapper and optimize illumination with the naive formulation (ablation) | Proposed · underwater scenes | optimization stability | Hurt |
Surprising findings
- Proposed-T scores slightly below SeaThru-NeRF in PSNR/SSIM on some SeaThru scenes such as JAPANESE in SSIM and REDSEA in SSIM (0.6240 vs 0.6446).
Limitations
- Water fluidity and fast currents are not modeled.
- Attenuation makes distant objects hard to recognize, destabilizing structure-from-motion camera poses.
- Dynamic branch relies on optical-flow-based motion masks from a pre-trained model.
Experimental setup
Datasets: SeaThru-NeRF dataset proposed in-the-wild underwater video dataset LLFF dataset (for environment transfer)
Benchmarks: SeaThru-NeRF dataset proposed underwater video dataset
Metrics: PSNR SSIM LPIPS user rating (realism, fidelity, MVC)
Baselines: SeaThru-NeRF Instant-NGP MIP-360 DynamicNeRF
Builds on: NeRF DynamicNeRF SeaThru-NeRF Instant-NGP RAFT-style optical flow
Closest papers by meaning
- Decoupling Scattering: Pseudo-Label Guided NeRF for Scenes with Scattering Media (AAAI 2025 · similarity 0.89)
- I2-NeRF: Learning Neural Radiance Fields Under Physically-Grounded Media Interactions (NeurIPS 2025 · similarity 0.87)
- A Pre-convolved Representation for Plug-and-Play Neural Illumination Fields (AAAI 2024 · similarity 0.85)
- Plenodium: Underwater 3D Scene Reconstruction with Plenoptic Medium Representation (NeurIPS 2025 · similarity 0.85)
- Neural NeRF Compression (ICML 2024 · similarity 0.84)
- Atlantis: Enabling Underwater Depth Estimation with Stable Diffusion (CVPR 2024 · similarity 0.84)
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.