UW-3DGS: Underwater 3D Reconstruction with Physics-Aware Gaussian Splatting

Changting Lin; Chaochao Chen; Jianfeng Dong; Jie Chen; Meng Han; Wenpeng Xing; Xun Zhou; Zaifeng Yang

REVIEW 4 major objections 3 minor 2 cited by

UW-3DGS: Underwater 3D Reconstruction with Physics-Aware Gaussian Splatting

T0 review · 4 major / 3 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims that coupling 3D Gaussian Splatting with a learnable underwater image formation model and uncertainty-based pruning yields state-of-the-art haze-free underwater reconstruction at 3DGS-level efficiency.

desk verdict A plausible 3DGS-based underwater reconstruction method with a real identifiability problem in its physics claims; worth refereeing if the full paper justifies the decomposition. read the letter →

arxiv 2508.06169 v1 pith:OVHTV56B submitted 2025-08-08 cs.CV cs.AI

Wenpeng Xing , Jie Chen , Zaifeng Yang , Changting Lin , Jianfeng Dong , Chaochao Chen , Xun Zhou , Meng Han This is my paper

classification cs.CVcs.AI

keywords underwater3DreconstructionGaussianSplattingimageformationmodelattenuationbackscatteruncertaintypruningfloatingartifacts

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to show that underwater 3D reconstruction can be made practical by embedding a learnable physics-based image formation model inside 3D Gaussian Splatting, rather than relying on the slow MLP regressions used by NeRF-based underwater methods. The model decomposes each hazy image into clean scene radiance, spatially varying attenuation, and backscatter, while a second branch prunes unstable floating Gaussians using uncertainty scores. On the SeaThru-NeRF and UWBundle benchmarks, the method reports PSNR 27.604, SSIM 0.868, LPIPS 0.104, and about 65 percent fewer floating artifacts. A sympathetic reader should take this as a claim that the physics can be learned at the same time as the geometry, and that clean haze-free views and physically re-rendered underwater views come from the same pipeline.

What carries the argument

The load-bearing object is the learnable underwater image formation module: it predicts spatially varying attenuation $\beta$ and backscatter $B$ on a voxel grid and applies the transmission model $T = \exp(-\beta d)$ along each view ray to turn clean radiance into a hazy image. This is what lets the Gaussian splatting optimizer separate what the scene looks like without water from what the water does to the light. The second mechanism, Physics-Aware Uncertainty Pruning, scores each Gaussian's uncertainty and removes unstable floaters, giving the geometry a way to become artifact-free instead of fitting noise.

What would settle it

Place a calibrated target at known depths in water with independently measured attenuation and backscatter, train the pipeline, and compare the recovered voxel maps of $\beta$ and $B$ to the measured values. If the recovered optical parameters are far off while the rendered images still look right, the physics-aware separation is a fitting artifact rather than a real light-transport estimate.

Watch

Extended reading notes

Core claim

The central claim is that 3D Gaussian Splatting, extended with a plug-and-play underwater image formation module, can jointly recover a clean scene and the optical properties of the water from ordinary training views. The formation module uses voxel-based regression to learn spatially varying attenuation and backscatter, so the classic formation equation $I = J T + B$ is inverted rather than assumed constant. The Physics-Aware Uncertainty Pruning branch removes noisy floating Gaussians, suppressing the artifacts that plague free-space reconstruction. Because the same Gaussian representation is rendered with and without media effects, the pipeline outputs both Unattenuated Radiance Images and

Load-bearing premise

The load-bearing premise is that the hazy training images are enough to pin down which part of the signal is the clean scene and which part is attenuation and backscatter, even though many different splits of the observed light can produce the same images.

Editorial extensions

If this is right

The same pipeline produces both media-free radiance images and physically rendered underwater views from one set of Gaussians.
If the claim holds, underwater reconstruction no longer has to choose between NeRF-quality physics and 3DGS efficiency; the two are combined in one end-to-end training stage.
The plug-and-play formation module can be attached to other 3DGS-style scene representations that face scattering media.
The PAUP pruning criterion offers a general way to suppress floating artifacts when optimizing free-space 3D Gaussians.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The paper does not demonstrate that the recovered attenuation and backscatter match independently measured water properties; the clean URIs should be treated as plausible reconstructions until such a comparison is made.
Because the formation equation is underdetermined, the reconstruction losses and pruning must be doing implicit regularization; ablating the formation module while keeping the same total model capacity would test whether the physics term is the source of the gains.
The same voxel-based media regression could be transferred to fog, smoke, or tissue imaging, where the same transmission-plus-backscatter equation holds.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

I'll be straight: this is a reasonable paper that deserves a careful look, but the headline physics claim is the weakness. What's new is the combination--3DGS with a voxel-based learnable attenuation/backscatter field and an uncertainty pruning branch to remove floating Gaussians. That's a natural step beyond SeaThru-NeRF, and the pruning idea is likely to help in practice. If the reported numbers hold, it's a useful engineering result for underwater scene capture.

The soft spot is identifiability. The formation model I = J*T + B*(1-T) is underdetermined from images alone: for any T and B you can define J from the residual. The stress-test note is right that without regularization or external validation of the recovered media parameters, the 'clean' URIs and 'accurate light transport' language could be fitting artifacts. This is the key issue for review. I don't think it's automatically fatal--multi-view consistency can pin down the decomposition in some cases, and the authors may have a smoothness prior or a physical constraint in the full paper. But the abstract gives no hint of it, and the full text I received was too garbled to check. If the full paper includes synthetic experiments with known ground-truth attenuation and backscatter, or a calibration against measured water properties, that would resolve it.

Minor things: the reported PSNR/SSIM/LPIPS are single numbers with no error bars, and there's no visible code or artifact link. Those are fixable in revision, not fatal.

Overall, I'd send this to peer review. The identifiability question is precisely what a good referee should pressure-test, and a conditional accept with a request for regularization details and synthetic validation would be fair. If the authors can show the decomposition is grounded, the paper is a solid contribution to degraded-media 3D vision.

Referee Report

4 major / 3 minor

Summary. UW-3DGS adapts 3D Gaussian Splatting to underwater 3D reconstruction by coupling a voxel-based learnable image formation module, which predicts spatially varying attenuation and backscatter from training views, with a Physics-Aware Uncertainty Pruning (PAUP) branch that removes noisy Gaussians. Training optimizes Gaussians and media parameters end-to-end; rendering produces clean unattenuated radiance images and underwater images with explicit light transport. On SeaThru-NeRF and UWBundle, the abstract reports PSNR 27.604, SSIM 0.868, LPIPS 0.104 and roughly 65% fewer floating artifacts.

Significance. The paper addresses a real problem: underwater images suffer wavelength-dependent absorption and scattering, and existing NeRF-based physics models are slow. Adapting 3DGS with a light transport model is a promising direction, and the PAUP idea is useful. If the recovered attenuation/backscatter maps are physically validated, this would be a solid contribution with practical efficiency gains. The main uncertainty is whether the decomposition is identifiable from the training data alone; the manuscript as supplied does not establish this.

major comments (4)

[Underwater image formation model (first equation, I = J·T + B·(1−T))] The central physics claim is not supported by the presented validation. Attenuation and backscatter are learned as per-voxel fields from the same training images used to fit the scene; for any chosen T and B, J is algebraically determined. Nothing in the available text ties the recovered T and B to measured water properties or synthetic ground truth. I therefore cannot distinguish the clean URI from a residual after fitting a flexible degradation model. Please add synthetic experiments with known β and B, or compare recovered β to independent measurements; otherwise soften the 'accurate light transport' claim.
[Abstract and quantitative evaluation] The reported results are single-point metrics with no variance, no number of runs, and no seed information. 3DGS training is stochastic, and pruning-based methods are sensitive to initialization and hyperparameters. Report mean ± std over multiple random seeds for each dataset and sequence. In addition, the claimed ~65% reduction in floating artifacts has no measurement protocol: define what counts as a floating artifact and how it is counted.
[Ablation study] The two key innovations are the learnable formation module and PAUP. The abstract reports only final-system numbers. To support the claim that these components are load-bearing, include ablations that (a) remove PAUP, (b) replace the voxel-based attenuation/backscatter with a global or per-sequence media model, and (c) use standard 3DGS without the formation module. If such ablations exist in the full text, they are not legible in the supplied version and need to be clearly presented.
[Manuscript integrity] The supplied manuscript text is heavily corrupted with mojibake; large sections, including equations and tables, are unreadable. This prevents verification of essential details such as the exact loss functions, hyperparameters, and experimental results. A clean, properly encoded version is required before the technical claims can be fully evaluated.

minor comments (3)

[Throughout] The notation J, T, B, and I is used without a consolidated definition near the abstract. Add a small notation table or clearly define these symbols at first use in the main text.
[Figures] Qualitative comparisons would benefit from error maps (e.g., absolute difference or LPIPS maps) and side-by-side views of recovered attenuation/backscatter versus input images, so readers can visually assess the media decomposition.
[PDF header] The PDF header metadata contains an unrelated arXiv identifier (2508.06173v1, cond-mat.mes-hall) and date. This metadata should be corrected to match the paper.

Circularity Check

1 steps flagged · score 4.0 of 10

Partial circularity: the 'clean' URIs and 'accurate light transport' are the residual and fitted values of the same learnable formation model, while the benchmark numbers remain independent held-out evaluations.

fitted input called prediction [Abstract (Key innovations (1), rendering stage)]
"Key innovations include: (1) a plug-and-play learnable underwater image formation module using voxel-based regression for spatially varying attenuation and backscatter; ... In rendering, refined Gaussians produce clean Unattenuated Radiance Images (URIs) free from media effects, while learned physics enable realistic Underwater Images (UWIs) with accurate light transport."

The underwater image formation model is of the form I = J*T + B*(1-T). The paper fits T (attenuation/transmission) and B (backscatter) on the training views with a voxel-regression module, and the clean URI J is then obtained as the algebraic residual J = (I - B*(1-T))/T. By construction, any fitted T and B reproduce the observed I and define a J, so the 'clean URIs free from media effects' and 'accurate light transport' are not independent predictions; they are the residual and the fitted values. The paper supplies no external validation of the recovered attenuation/backscatter against measured water properties or synthetic ground truth, so the physical decomposition claim is partially forced by the model's own definition.

full rationale

The paper's strongest benchmark claims (PSNR 27.604, SSIM 0.868, LPIPS 0.104 on SeaThru-NeRF, ~65% floating-artifact reduction) are evaluated on held-out test views from external datasets, so those numbers do not reduce to the fitted parameters. The circular component is narrower: the learnable underwater formation module is trained on the same views that define the scene, and the clean URIs are derived from the same formation equation after subtracting the fitted backscatter and dividing by the fitted transmission. Thus the 'physics-aware' output is not an independently measured or validated quantity; it is the residual of a fitted decomposition. The engineering comparison remains meaningful, but the specific claim that the method recovers true unattenuated radiance and accurate light transport is partially circular unless identifiability is established by external constraints or synthetic validation. No load-bearing self-citations were identifiable in the supplied text.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

Everything hangs on the learned media maps: the paper fits attenuation and backscatter per voxel to the very images whose reconstruction it reports, so claims of physical accuracy inherit all the identifiability problems of the underwater formation equation. No new physical entities are introduced; the URI and UWI outputs are derived renderings, not independent degrees of freedom.

free parameters (3)

voxel-wise attenuation map (per channel) = unknown, learned during training
The learnable formation module regresses spatially varying attenuation per voxel from the training images; these are fitted values, not measured, and they drive the claimed physical decomposition.
voxel-wise backscatter map (per channel) = unknown, learned during training
Backscatter is regressed jointly with attenuation and scene radiance; since multiple (clean scene, media) triples fit the same observations, these values are not independently identified.
PAUP pruning threshold and uncertainty hyperparameters = not stated in abstract
The adaptive pruning branch must rely on a threshold or regularizer chosen by hand; the abstract does not report it, yet it controls the 65 percent artifact-reduction claim.

assumptions (3)

domain assumption Underwater image formation: observed color equals direct attenuated scene radiance plus scattered backscatter
Invoked by the abstract's learnable formation module; this is the standard SeaThru-type model, but its identifiability from images alone is assumed, not proven.
domain assumption Attenuation and backscatter vary smoothly enough that voxel-based regression yields a unique decomposition
Required to disambiguate clean radiance from media effects; the abstract states no regularization or prior that makes the split unique.
domain assumption 3D Gaussian Splatting differentiable rendering faithfully models scene appearance across views
The whole reconstruction rests on standard 3DGS machinery as background; it is not re-derived or validated inside this paper.

how reviews work

0 comments

Cite this review

Pith. "Pith review of UW-3DGS: Underwater 3D Reconstruction with Physics-Aware Gaussian Splatting." pith.science (2026). https://pith.science/paper/OVHTV56B

@misc{pith2026250806169,
  author       = {Pith},
  title        = {Pith review of: UW-3DGS: Underwater 3D Reconstruction with Physics-Aware Gaussian Splatting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OVHTV56B}},
  note         = {Machine review of arXiv:2508.06169}
}

read the original abstract

Underwater 3D scene reconstruction faces severe challenges from light absorption, scattering, and turbidity, which degrade geometry and color fidelity in traditional methods like Neural Radiance Fields (NeRF). While NeRF extensions such as SeaThru-NeRF incorporate physics-based models, their MLP reliance limits efficiency and spatial resolution in hazy environments. We introduce UW-3DGS, a novel framework adapting 3D Gaussian Splatting (3DGS) for robust underwater reconstruction. Key innovations include: (1) a plug-and-play learnable underwater image formation module using voxel-based regression for spatially varying attenuation and backscatter; and (2) a Physics-Aware Uncertainty Pruning (PAUP) branch that adaptively removes noisy floating Gaussians via uncertainty scoring, ensuring artifact-free geometry. The pipeline operates in training and rendering stages. During training, noisy Gaussians are optimized end-to-end with underwater parameters, guided by PAUP pruning and scattering modeling. In rendering, refined Gaussians produce clean Unattenuated Radiance Images (URIs) free from media effects, while learned physics enable realistic Underwater Images (UWIs) with accurate light transport. Experiments on SeaThru-NeRF and UWBundle datasets show superior performance, achieving PSNR of 27.604, SSIM of 0.868, and LPIPS of 0.104 on SeaThru-NeRF, with ~65% reduction in floating artifacts.

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

UR$^2$: Unify RAG and Reasoning through Reinforcement Learning
cs.CL 2025-08 unverdicted novelty 6.0 of 10

UR^2 trains LLMs with reinforcement learning to dynamically combine retrieval and reasoning, beating RAG and RL baselines across four domains.
UR$^2$: Unify RAG and Reasoning through Reinforcement Learning
cs.CL 2025-08 unverdicted novelty 5.0 of 10

UR² is a general RL framework that dynamically coordinates RAG and reasoning via difficulty-aware curriculum and hybrid knowledge access, outperforming baselines on QA, MMLU-Pro, medical, and math tasks with models up...

Reference graph

Works this paper leans on

1 extracted references · 1 canonical work pages · cited by 1 Pith paper

[1]

� �� ...

work page Pith review arXiv 2025

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

� �� ...

work page Pith review arXiv 2025