{"id":"1ba08a22-1529-45af-92f0-a1762de3d05a","arxiv_id":"2505.15737","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"RUSplatting improves sparse-view underwater 3D reconstruction by decoupling RGB attenuation, synthesizing intermediate frames, and adding edge-aware smoothness, with a new deep-sea dataset.","lead":"This paper adapts 3D Gaussian Splatting for deep underwater scenes by separating color channels, adding interpolated frames, and applying edge-aware smoothing, reporting better reconstruction quality than existing methods. It also releases a new deep-sea shipwreck dataset and code, so other groups can test the approach.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Pseudo-depth quality and per-scene evaluation gaps leave the 1.90 dB claim load-bearing on an untested assumption.","rationale":"The reader's weakest_assumption identifies the same central risk: the method's reliance on Depth-Anything-V2 pseudo-depth as the sole geometric supervision and as the depth input to the underwater attenuation model in deep-sea conditions. This is indeed the most load-bearing assumption because it threads through three different parts of the pipeline (depth supervision, ESL edge weights, and the physical transmission model), and the manuscript itself flags that the tool was not designed for underwater imagery. My added specificity is that this dependency is untested: there is no ablation or sensitivity analysis varying pseudo-depth quality, and Table 1 only reports dataset-level averages, so the headline 'up to 1.90 dB' is not tied to an error-barred, per-scene distribution. The hyperparameter-selection concern (α, λb, λs chosen on the evaluation data) is secondary but reinforces the need for a conditional verdict: the empirical gains are plausible and the artifact release is genuinely valuable, but the robustness claim needs a targeted depth-sensitivity check and per-scene reporting before full acceptance. I found no internal inconsistency in the main equations; Eq. 6 follows from Eq. 3, and the loss formulation is coherent. Thus the appropriate verdict remains CONDITIONAL, exactly as the reader proposed, with the concrete test being the pseudo-depth perturbation experiment.","tokens_in":10583,"tokens_out":2122,"duration_ms":17657,"concrete_test":"Run a pseudo-depth sensitivity experiment on the S-UW or Submerged3D test split: train RUSplatting with the official Depth-Anything-V2 pseudo-depth as reported; retrain with pseudo-depth deterministically degraded per scene—e.g. multiplicative scale perturbation (0.8–1.2) and additive low-frequency bias approximating turbidity falloff; retrain with per-scene optimal affine alignment of pseudo-depth to COLMAP sparse depth. If held-out PSNR/SSIM drops by more than ~0.5 dB under perturbation, or improves materially under affine correction, the depth assumption is load-bearing and the 'consistently outperforms' claim should be reported per scene with error bars.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central empirical claim is that RUSplatting consistently outperforms baselines with PSNR gains up to 1.90 dB (and 2.83–3.84 dB vs. best baselines). The weakest load-bearing assumption is that Depth-Anything-V2 pseudo-depths are accurate enough in deep-sea, low-light, high-turbidity conditions to serve simultaneously as (i) geometric supervision, (ii) edge weights in ESL (Eq. 9), and (iii) the depth z in the per-channel transmission model (Eq. 5). The manuscript itself concedes in Section 6 that RIFE and Depth-Anything-V2 are not designed for underwater conditions. Since LDepth directly enforces rendered depth against pseudo-depth and the attenuation decoupling uses the same z, systematic pseudo-depth errors—e.g. scale ambiguity or turbidity-dependent bias—do not cancel but propagate into both geometry and colour recovery. Table 1 reports only dataset-level averages, so a single strong scene could drive the reported gains; without per-scene numbers or error bars, the 1.90 dB headline and the 'consistently outperforms' wording cannot be verified as a robust across-scene effect. The absence of a sensitivity analysis over pseudo-depth quality leaves this dependency untested. A second, related gap is the lack of tuning transparency: hyperparameters α, λb, λs and the deeper-MLP choice are selected on evaluation data, so part of the gain may reflect favourable hyperparameter search rather than the architectural contributions. The concern is not that pseudo-depth is invalid in principle—Depth-Anything priors are common in sparse-view 3DGS—but that the specific claim of robustness in deep-sea conditions is not yet supported by evidence that the method tolerates the pseudo-depth errors likely there.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes RUSplatting, a 3D Gaussian Splatting framework for sparse-view underwater scene reconstruction. The method combines four contributions: per-RGB-channel decoupling of underwater attenuation and backscatter parameters estimated by a deeper MLP, an intermediate frame interpolation (IFI) mechanism with adaptive per-frame uncertainty weighting (AFW), an edge-aware smoothness loss (ESL) that uses pseudo-depth discontinuities to denoise while preserving structure, and a newly collected deep-sea dataset, Submerged3D. Experiments are reported on SeaThru-NeRF, S-UW, and Submerged3D against five baselines, with the claim that RUSplatting consistently outperforms all baselines across PSNR, SSIM, and LPIPS.","tokens_in":10958,"tokens_out":6710,"duration_ms":54856,"significance":"If the empirical claims hold, the paper offers a practical and well-motivated recipe for reconstructing sparse-view, low-light underwater scenes, and the release of a new deep-sea dataset and public code is a useful community contribution. The core priors (Depth-Anything-V2, RIFE, R-UIFM) are external and not produced by the authors, so circularity risk is low. However, the current evidence does not yet substantiate the headline claims: there are unresolved quantitative inconsistencies between the abstract, Section 5.2, Table 1, and the ablation study, and the reported robustness of the method is not supported by per-scene results, error bars, or a sensitivity analysis of the pseudo-depth dependency.","major_comments":[{"comment":"The abstract's claim of 'PSNR gains up to 1.90dB' is not supported by Table 1. The largest improvement over the best baseline in Table 1 is 1.22 dB (Submerged3D, RUSplatting 25.7990 vs. WaterSplatting 24.5832); the 2.83, 3.01, and 3.84 dB figures quoted in Section 5.2 correspond to the average improvement over all five baselines, not the improvement over the best baseline, and this should be stated explicitly. The authors should either reconcile the 1.90 dB figure with a specific scene-level result or remove it.","section":"Abstract and Section 5.2, Table 1"},{"comment":"Depth-Anything-V2 pseudo-depth is used simultaneously as the depth z in the per-channel transmission model (Eq. 5), as the supervision target for LDepth (Eq. 10), and as the edge-weight source in ESL (Eq. 9). Because Section 6 concedes that Depth-Anything-V2 is not designed for underwater conditions, systematic errors in pseudo-depth could propagate into both the attenuation decoupling and the geometry supervision. The paper provides no sensitivity analysis, no comparison with alternative depth sources, and no depth-quality diagnostics. A concrete test, such as perturbing pseudo-depth or evaluating on a scene with known depth, is needed to establish that the reported gains are robust to this dependency.","section":"Section 4.3, Section 4.4, Section 5.1"},{"comment":"The claim that RUSplatting 'consistently outperforms' all baselines across all metrics is based solely on dataset-level averages over four scenes per dataset. No per-scene numbers, standard deviations, or significance tests are reported. This is particularly important because the Submerged3D SSIM margin over UW-GS is only 0.0001 (0.7724 vs. 0.7723), and the LPIPS differences are small. Per-scene tables and error bars are needed to verify the across-scene robustness that the word 'consistently' implies.","section":"Table 1 and Section 5.2"},{"comment":"The ablation study contains a numerical inconsistency with Table 1. The text reports that the full RUSplatting model achieves an average PSNR of 25.7874 across all datasets, but the mean of the three dataset-averaged PSNR values in Table 1 is 26.8958. The ablation also reports M5 at 25.03 PSNR, implying an AFW gain of 0.76 dB, yet no per-dataset ablation breakdown is given. The authors should clarify the exact subset and weighting used in the ablation averages and provide per-scene ablation results.","section":"Section 5.3"},{"comment":"Hyperparameters such as alpha, lambda_b, lambda_s, the loss weights, and the choice of a 5-layer MLP are selected via grid search or hyperparameter sensitivity analysis, but no validation split is described. If these choices are tuned on the test scenes used for evaluation, part of the reported improvement could reflect favorable hyperparameter selection rather than the proposed components. The validation protocol should be stated, and a fixed validation split or nested selection procedure should be used.","section":"Sections 4.2–4.3 and Section 5.1"}],"minor_comments":[{"comment":"The abstract claims improved 'geometric accuracy,' but no quantitative geometric evaluation is reported; the paper relies on pseudo-depth only as weak supervision. Either add depth evaluation metrics or temper the geometric-accuracy claim.","section":"Abstract and Section 1"},{"comment":"The ablation configuration table marks both 'IFI w/o AFW' and 'IFI w/ AFW' as enabled for the final RUSplatting column, which is confusing. The rows should clearly indicate that the final model uses IFI with AFW only.","section":"Section 5.3, Figure 4"},{"comment":"The gamma notation is inconsistent: Eq. (7) uses gamma_f for each interpolated frame, while Eq. (11) writes only gamma. The subscript should be restored in Eq. (11) for clarity.","section":"Eq. (7) and Eq. (11)"},{"comment":"The listed percentage improvements '21.37%, 4.99%, and 5.09%' relative to SeeThru-NeRF, WaterSplatting, and UW-GS do not have their baseline definitions stated; specify whether these are relative PSNR improvements and against which exact baseline configurations.","section":"Section 1"},{"comment":"The table caption says red bold denotes the best result and yellow the second best, but no yellow highlighting is visible in the provided text; ensure the final rendering shows both annotations or revise the caption.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":"The paper's own prior work (UW-GS and the review paper [10]) involves overlapping authors with this submission, and the strongest baseline numbers happen to be from those same-group methods. The manuscript uses official implementations for comparisons, so this is not a circularity problem, but an independent evaluation of the Submerged3D dataset would strengthen confidence. The unresolved numerical discrepancy between the abstract's 1.90 dB claim and Table 1 should be fixed before the paper can be considered for acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a careful look. The genuinely new pieces are the RGB-decoupled medium estimation with a deeper MLP, the interpolated-frame pipeline with uncertainty-based adaptive weighting, the edge-aware smoothness loss, and the Submerged3D dataset of deep-sea wrecks. Code and data are public, which is more than many papers in this niche provide. The ablation study is cleanly structured and each component earns its keep; the IFI and AFW ablations in particular show large, plausible effects.\n\nThe soft spots are the ones the reader flagged. Table 1 reports only dataset-level averages, and the text claims gains over baselines that don't match the abstract's 1.90 dB figure — the table shows 2.8–3.8 dB over the best baseline, so the abstract actually undersells it, but the lack of per-scene numbers and error bars makes \"consistently outperforms\" hard to verify. Hyperparameters (λs, λb, α, deeper-MLP choice) are selected via grid search without a clear validation split, so some of the gain may be tuning. More fundamentally, pseudo-depth from Depth-Anything-V2 is used in three load-bearing places: depth supervision, edge weights in ESL, and the z in the per-channel transmission model. The authors themselves concede in Section 6 that RIFE and Depth-Anything-V2 aren't designed for underwater. If pseudo-depth is biased in deep-sea turbidity, the attenuation decoupling and depth loss push in the same wrong direction. That doesn't sink the paper — the method still works better than baselines in the reported conditions — but it means the robustness claim is narrower than the title suggests. A sensitivity analysis on pseudo-depth quality (e.g., scaling or perturbing D) would address this directly.\n\nOn balance, this is a legitimate extension-level contribution with a useful dataset and public code. The central empirical claim is plausible, not airtight. I'd send it to peer review and ask for per-scene results, error bars, and a pseudo-depth sensitivity study in the revision. The stress-test note overstates the 1.90 dB issue slightly — that number appears to come from a different comparison — but the underlying concern about verification is fair.","headline":"Solid specialized contribution with a new deep-sea dataset, but the headline PSNR gain is not backed by per-scene statistics and the method leans heavily on pseudo-depths that the authors admit are off-domain.","tokens_in":11479,"tokens_out":1795,"would_cite":true,"duration_ms":15692,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"RUSplatting claims sparse-view underwater 3D reconstruction improves by coupling per-channel water-attenuation physics into Gaussian splatting, yielding PSNR gains up to 1.90 dB.","keywords":["underwater 3D reconstruction","3D Gaussian splatting","sparse-view novel view synthesis","underwater image formation model","frame interpolation","adaptive weighting","edge-aware smoothness loss","deep-sea dataset"],"falsifier":"Train RUSplatting on an underwater scene where ground-truth depth is available (for example, a synthetic water tank or a capture with a depth sensor) and compare against the pseudo-depth variant; if replacing Depth-Anything-V2 depths with true depths does not improve or preserves the reported gains, the claimed coupling between depth accuracy and medium estimation is not supported.","tokens_in":10437,"feed_emoji":"🌊","tokens_out":5711,"duration_ms":46173,"temperature":0.7,"pith_summary":"This paper tries to establish that 3D Gaussian Splatting can reconstruct deep underwater scenes from sparse views if the water's wavelength-dependent absorption and scattering are modeled per color channel and if training is enriched with interpolated frames and edge-aware noise suppression. This matters because marine archaeology, biology, and robotics rely on reconstructing sites from limited imagery, where existing methods assume clear air or dense, temporally coherent input. The authors claim consistent gains over state-of-the-art baselines on three datasets, including their new deep-sea Submerged3D set, with best PSNR gain of 1.90 dB.","feed_headline":"Sparse-view underwater 3D reconstruction gains up to 1.90 dB","feed_subtitle":"Splitting color channels by water physics improves deep-sea 3D rendering from sparse views.","key_machinery":"The central mechanism is per-channel medium-parameter prediction: an MLP takes position-encoded depth and viewpoint and outputs medium parameters for each Gaussian and each color channel, producing the corrected color $c^m_{i,\\text{ch}} = T^D_{i,\\text{ch}} \\cdot c_{i,\\text{ch}} + (1 - T^B_{i,\\text{ch}}) \\cdot b_{i,\\text{ch}}$. Supporting this, Intermediate Frame Interpolation (IFI) uses RIFE to synthesize frames between sparse inputs, Adaptive Frame Weighting (AFW) tunes each interpolated frame's loss weight via a learnable uncertainty parameter, and an Edge-Aware Smoothness Loss (ESL) uses depth discontinuities from pseudo-depth maps as edge weights to smooth noise while preserving structure.","core_discovery":"On the paper's own terms, the central discovery is that decoupling the underwater image formation physics per RGB channel, with attenuation $T^D_{\\text{ch}}$ and backscatter $T^B_{\\text{ch}}$ varying with distance $z$, inside an affine color transform of each Gaussian, combined with interpolated intermediate frames and a depth-guided edge-aware smoothness loss, makes sparse-view underwater novel-view synthesis robust. The method builds directly on 3D Gaussian Splatting and replaces a shallow MLP with a five-layer MLP that predicts per-channel medium parameters $\\{T^D, T^B, \\beta^d, \\beta^b, b\\}$. The authors report that RUSplatting outperforms five baselines across PSNR, SSIM, and LPIPS on SeaThru-NeRF, S-UW, and Submerged3D.","pith_inferences":["If the per-channel physics decoupling is the true source of gain, the same affine model could transfer to other turbid media such as fog or murky rivers, where wavelength-dependent attenuation is weaker; a direct comparison on fog data would test that transfer.","The method's reliance on Depth-Anything-V2 and RIFE suggests that fine-tuning these estimators on underwater imagery, or substituting physics-based depth estimates, would further improve robustness — a limitation the authors themselves note.","A stronger diagnostic would be rendering only the training views rather than held-out views, to see whether the PSNR gain reflects improved geometry or mainly improved color post-processing.","The depth-guiding role of ESL means the method's noise suppression quality is coupled to monocular depth accuracy; comparing against a variant using edge weights from the rendered image itself would isolate that dependency."],"forward_implications":["Underwater scenes with roughly 20 training views per sequence can be rendered at higher fidelity than previous underwater NeRF and Gaussian Splatting methods, supporting interactive onshore inspection of shipwrecks and marine sites.","Per-channel affine color correction restores color in high-turbidity deep-sea footage without a separate image-enhancement network, reducing color cast during reconstruction itself.","Interpolated frames combined with learnable weighting make the pipeline more tolerant of wide baselines and limited overlap between input views.","The release of Submerged3D provides a real deep-sea benchmark with low light, severe turbidity, and shipwreck scenes, enabling direct comparisons for future underwater reconstruction work."],"supporting_citations":[{"why":"Supplies the revised underwater image formation model with per-channel attenuation and backscatter equations that underpin the decoupled color correction.","marker":"[1]"},{"why":"Provides the base 3D Gaussian Splatting representation and alpha-blending rendering pipeline that RUSplatting extends.","marker":"[15]"},{"why":"Provides the SeaThru-NeRF baseline and one of the evaluation datasets, along with the idea of embedding water physics in neural rendering.","marker":"[17]"},{"why":"Serves as a WaterSplatting baseline for underwater Gaussian Splatting with affine medium transforms and as a comparison target.","marker":"[18]"},{"why":"Supplies the UW-GS baseline, the binary motion mask used to exclude distractors, the S-UW dataset, and prior physics-driven density control.","marker":"[27]"},{"why":"Produces the Depth-Anything-V2 pseudo-depth maps used both for depth supervision and for the depth-discontinuity edge weights in the smoothness loss.","marker":"[31]"},{"why":"RIFE is the frame-interpolation network used to generate intermediate frames for the IFI component.","marker":"[11]"},{"why":"Motivates the uncertainty-based adaptive weighting scheme used to down-weight poorly interpolated frames.","marker":"[13]"}],"fun_headline_variants":["Physics-guided color separation improves sparse-view underwater 3D","Sparse-view deep-sea 3D reconstruction gains 1.90 dB PSNR","New Submerged3D dataset powers robust underwater Gaussian splatting","RUSplatting: edge-preserving sparse-view underwater 3D with physics"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The single load-bearing assumption is that pseudo-depth maps from Depth-Anything-V2 are accurate enough in deep-sea, low-light, high-turbidity images to supervise geometry and to set edge weights in the smoothness loss; if they are not, the decoupled physics and depth losses push the optimization with incorrect depth values.","fun_headline_variants_meta":{"raw":{"variants":["Physics-guided color separation improves sparse-view underwater 3D","Sparse-view deep-sea 3D reconstruction gains 1.90 dB PSNR","New Submerged3D dataset powers robust underwater Gaussian splatting","RUSplatting: edge-preserving sparse-view underwater 3D with physics"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000328,"raw_usage":{"total_tokens":1829,"prompt_tokens":936,"completion_tokens":893,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":552,"completion_tokens_details":{"reasoning_tokens":812}},"tokens_in":552,"tokens_out":893,"duration_ms":6719,"temperature":1.0,"reasoning_tokens":812,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T15:11:52.853154+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train RUSplatting on an underwater scene where ground-truth depth is available (for example, a synthetic water tank or a capture with a depth sensor) and compare against the pseudo-depth variant; if replacing Depth-Anything-V2 depths with true depths does not improve or preserves the reported gains, the claimed coupling between depth accuracy and medium estimation is not supported.","supporting_citations":[{"cited_title":"A revised underwater image formation model","cited_arxiv_id":null,"evidence_quote":"Supplies the revised underwater image formation model with per-channel attenuation and backscatter equations that underpin the decoupled color correction."},{"cited_title":"Seathru-nerf: Neural radiance fields in scattering media","cited_arxiv_id":null,"evidence_quote":"Provides the SeaThru-NeRF baseline and one of the evaluation datasets, along with the idea of embedding water physics in neural rendering."},{"cited_title":"Wa- terSplatting: Fast underwater 3D scene reconstruction using gaussian splatting","cited_arxiv_id":null,"evidence_quote":"Serves as a WaterSplatting baseline for underwater Gaussian Splatting with affine medium transforms and as a comparison target."},{"cited_title":"UW-GS: Distractor-aware 3D gaussian splatting for enhanced underwater scene reconstruction","cited_arxiv_id":null,"evidence_quote":"Supplies the UW-GS baseline, the binary motion mask used to exclude distractors, the S-UW dataset, and prior physics-driven density control."},{"cited_title":"Real-time intermediate flow estimation for video frame interpolation","cited_arxiv_id":null,"evidence_quote":"RIFE is the frame-interpolation network used to generate intermediate frames for the IFI component."}],"review_version":1}