{"id":"637ba771-ee95-4a3b-99be-b453ab9f7f58","arxiv_id":"2508.06169","paper_version":1,"verdict":"CONDITIONAL","confidence":"LOW","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"UW-3DGS adapts 3D Gaussian Splatting to underwater scenes with a voxel-based learnable attenuation/backscatter model and uncertainty pruning, reporting higher PSNR/SSIM and lower LPIPS than SeaThru-NeRF.","lead":"UW-3DGS rebuilds 3D scenes from underwater photos using Gaussian Splatting, with a learned model of water absorption and scattering and a pruning step that removes floating ghost artifacts. It matters because faster, cleaner underwater reconstruction helps ocean inspection, marine archaeology, and robotic diving.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"If attenuation/backscatter are identifiable only through reconstruction losses, the 'clean' URIs and accurate light transport are fitting artifacts; synthetic ground-truth validation is needed.","rationale":"The reader's verdict is CONDITIONAL and its weakest assumption is precisely the identifiability of the clean-versus-media decomposition. I agree that this is the load-bearing point: the paper's advertised novelty is a learnable physics-aware formation module, and the superiority claim is tied to recovering physically meaningful parameters. The concern is not an ad hominem or a matter of consensus; it is an internal identifiability risk in the stated formation model. The proposed synthetic experiment would settle it: with ground-truth media parameters and clean radiance known, either the method recovers them or it does not. If it does not, the authors would need to either add constraints (e.g., priors on beta, known water type, or explicit validation against measured backscatter) or soften the physics claims. Error bars and code availability, while useful, are secondary and do not change the merit determination. Because the concern is real but testable, and because the reader already issued a conditional verdict, my recommendation is to keep the verdict unchanged rather than escalate or downgrade.","tokens_in":18439,"tokens_out":3318,"duration_ms":39165,"concrete_test":"Create a synthetic underwater scene with known clean radiance J_gt, spatially varying attenuation beta_gt, and backscatter B_gt. Render a set of training views with the same formation model, train UW-3DGS, and compare the recovered J_est, beta_est, B_est to ground truth using normalized RMSE for beta and B and PSNR/SSIM/LPIPS for the clean URI. If the recovered parameters and clean radiance match ground truth within acceptable tolerance, identifiability is empirically supported. Also run an ablation that replaces the learnable formation module with a per-pixel learnt color correction while keeping PAUP; if the resulting PSNR difference is small (e.g., below 0.5 dB), the reported gains are not attributable to the physics model.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central physics claim rests on decomposing each observed underwater image into clean scene radiance J, transmission T = exp(-int beta dz), and backscatter B via I = J*T + B*(1-T). From training views alone, this decomposition is highly underdetermined: for any chosen beta and B, one can algebraically define J = (I - B*(1-T))/T, yielding an infinite family of (clean scene, media) pairs that reproduce the observed images exactly. The voxel-based regression plus reconstruction losses may therefore select an arbitrary member of this family, not the physically correct one. Nothing in the abstract ties the recovered beta or B to measured water properties, and the supplied full text is too garbled to verify whether a regularizer or external constraint exists. If the decomposition is arbitrary, the reported PSNR 27.604 / SSIM 0.868 / LPIPS 0.104 gains and the 65% floating-artifact reduction could be produced by model capacity and pruning alone, while the 'physics-aware' and 'accurate light transport' claims become unfalsified decoration. The engineering result would remain interesting, but the paper's key conceptual contribution would be weakened to a purely empirical fitting method.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":18651,"tokens_out":5239,"duration_ms":64530,"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":[{"comment":"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.","section":"Underwater image formation model (first equation, I = J·T + B·(1−T))"},{"comment":"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.","section":"Abstract and quantitative evaluation"},{"comment":"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.","section":"Ablation study"},{"comment":"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.","section":"Manuscript integrity"}],"minor_comments":[{"comment":"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.","section":"Throughout"},{"comment":"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.","section":"Figures"},{"comment":"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.","section":"PDF header"}],"recommendation":"major_revision","confidential_remarks":"The paper has a plausible contribution, but the current submission is not reviewable in its present form because of the text corruption and because the central physics claim lacks the validation described in my major comments. I recommend requesting a clean resubmission with synthetic-ground-truth media evaluation, variance-aware metrics, and ablations before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","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.\n\nThe 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.\n\nMinor 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.\n\nOverall, 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.","headline":"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.","tokens_in":19240,"tokens_out":2819,"would_cite":false,"duration_ms":33348,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"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.","keywords":["underwater 3D reconstruction","3D Gaussian Splatting","image formation model","attenuation","backscatter","uncertainty pruning","floating artifacts"],"falsifier":"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.","tokens_in":18262,"feed_emoji":"🌊","tokens_out":5584,"duration_ms":61481,"temperature":0.7,"pith_summary":"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.","feed_headline":"Physics-aware Gaussian splatting clears underwater haze in 3D","feed_subtitle":"New method learns attenuation and backscatter while pruning floaters, hitting PSNR 27.6 on SeaThru-NeRF.","key_machinery":"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.","core_discovery":"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","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[],"fun_headline_variants":["Splatting learns water physics to build clear 3D scenes","Underwater 3D: Gaussian splatting that sees through haze","Physics-aware splatting prunes floaters, reveals clean underwater 3D","Inverting water optics with Gaussian splatting for artifact-free 3D","New splatting cuts underwater floaters, boosts 3D clarity"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Splatting learns water physics to build clear 3D scenes","Underwater 3D: Gaussian splatting that sees through haze","Physics-aware splatting prunes floaters, reveals clean underwater 3D","Inverting water optics with Gaussian splatting for artifact-free 3D","New splatting cuts underwater floaters, boosts 3D clarity"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000306,"raw_usage":{"total_tokens":1618,"prompt_tokens":797,"completion_tokens":821,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":541,"completion_tokens_details":{"reasoning_tokens":723}},"tokens_in":541,"tokens_out":821,"duration_ms":7986,"temperature":1.0,"reasoning_tokens":723,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T22:53:49.433482+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[],"review_version":1}