{"id":"a7b89f84-7d54-41ee-b831-341cf74a05ba","arxiv_id":"2505.01869","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":2.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A structured review of underwater visual enhancement and 3D reconstruction methods, from classical physics-based approaches to NeRF and 3D Gaussian Splatting, with a small qualitative comparison.","lead":"This paper reviews methods for fixing degraded underwater images and for turning them into 3D models, bringing together classical physics-based approaches and modern neural rendering. It is most useful as an orientation guide for researchers and engineers entering underwater computer vision.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section 5 does not deliver the quantitative evaluation promised in the abstract; the ranking of UW-GS as best rests on visual inspection of two scenes using the authors' own methods, without reproducibility.","rationale":"The reader's verdict (CONDITIONAL) already flags the benchmark's reliance on visual inspection and the authors' own unpublished enhancement. My stress-test sharpens this into a more concrete internal inconsistency: the abstract claims quantitative evaluation, but Section 5 contains no numbers. This mismatch directly affects the paper's stated contributions. The survey content itself is broad, well-structured, and consistent with the cited literature; the physics background and taxonomy are useful and credible. The benchmark section, however, is presented as a contribution and its central conclusion (UW-GS best) is not supported by the evidence shown. Because the deficiency is localized to the benchmark and fixable by adding proper quantitative evaluation and releasing code, a conditional-accept verdict remains appropriate rather than rejection. The paper should clarify that the benchmark is illustrative and either remove the 'quantitative' claim from the abstract or supply the missing metrics.","tokens_in":44931,"tokens_out":4814,"duration_ms":47017,"concrete_test":"Re-run the Section 5.3 comparison on all scenes of the SeaThru and S-UW datasets (not just Panama and Reef), computing PSNR, SSIM, and LPIPS on held-out views for Instant-NGP, SeaThru-NeRF, 3DGS, and UW-GS over at least five random seeds. Require the authors to release the UW-GS code and the Huang et al. (2025) enhancement code, or replace them with publicly available implementations. If the numerical aggregate does not place UW-GS first on a majority of scenes, or if rankings vary across seeds, the central empirical claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract states that the paper conducts 'both quantitative and qualitative evaluations of state-of-the-art UVE and underwater 3D reconstruction algorithms across multiple benchmark datasets,' yet Section 5 reports no quantitative metrics. Figure 19 compares Instant-NGP, SeaThru-NeRF, 3DGS, and UW-GS on only two scenes (Panama from SeaThru, Reef from S-UW) and concludes 'UW-GS appears to be the best' purely from visual inspection. No PSNR, SSIM, LPIPS, error bars, seeds, or released code are provided. The comparison is further compromised because UW-GS is the authors' own method (Wang et al., 2025, WACV) and the two-stage pipeline in Section 5.2 uses their unpublished enhancement method (Huang et al., 2025), making the results unreproducible by independent researchers. The paper itself concedes that public underwater 3D datasets are rare (Table 7; Section 5), and the limited scenes are not necessarily representative of the diversity of underwater conditions. If the benchmark is intended as a substantive contribution, the evidence is insufficient; if it is merely illustrative, the abstract's quantitative claim should be retracted.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript presents a review of underwater visual enhancement (UVE) and underwater 3D reconstruction. It introduces the Jaffe–McGlamery model and simplified image formation models, organizes UVE methods into conventional, data-driven, and hybrid categories, and reviews reconstruction approaches ranging from photogrammetry and visual SLAM to NeRF and 3D Gaussian Splatting. The paper also reports a small benchmark of three pipelines: reconstruction without enhancement, two-stage enhancement-then-reconstruction, and integrated physics-based reconstruction, with a qualitative comparison on two scenes (Panama and Reef). The authors claim this is the first unified review covering both UVE and underwater 3D reconstruction.","tokens_in":45143,"tokens_out":4691,"duration_ms":43432,"significance":"If the promised quantitative evaluation were delivered, the benchmark would provide a useful reference point for practitioners. The paper's taxonomy is broad and current, including Mamba- and diffusion-based enhancement methods and recent underwater Gaussian splatting approaches, and it usefully compiles public datasets and discusses open challenges. However, the central empirical contribution currently rests on qualitative inspection of two scenes and on methods with overlapping authorship, which limits the independent value of the benchmark.","major_comments":[{"comment":"The abstract states that the paper conducts 'both quantitative and qualitative evaluations' of state-of-the-art UVE and underwater 3D reconstruction algorithms across multiple benchmark datasets, but Section 5 reports no quantitative metrics (no PSNR, SSIM, LPIPS, or error bars), and the conclusion that 'UW-GS appears to be the best' is drawn from visual inspection of two scenes. This gap bears directly on contribution (3) in Section 1.3. Please either add a quantitative protocol with per-scene metrics and multiple random seeds, or revise the abstract and Section 1.3 to describe Section 5 as a qualitative illustration only.","section":"Section 5 and Abstract; Figure 19"},{"comment":"The two-stage pipeline uses the authors' unpublished enhancement method (Huang et al., 2025), and the integrated pipeline highlights UW-GS (Wang et al., 2025), which shares authors with this survey. No code, checkpoints, or detailed hyperparameters are provided, so an independent researcher cannot reproduce or verify the comparison. Please make the benchmark reproducible and include at least one independent baseline not affiliated with the authors, or clearly frame these results as self-reported demonstrations.","section":"Sections 5.2 and 5.3; Huang et al. (2025) and Wang et al. (2025)"},{"comment":"The claim that 'a comprehensive and systematic review covering both UVE and underwater 3D reconstruction remains absent' is not supported by a comparison with existing surveys, such as Anwar and Li (2020) for underwater image enhancement or Diamanti and Ødegård (2024) for marine 3D documentation. Please add a survey-comparison table or qualify the novelty claim.","section":"Section 1.3 and Abstract"},{"comment":"Table 7 shows that public underwater 3D datasets are scarce and small, and the headline comparison in Figure 19 uses only two scenes. The paper's own observation that current public data are limited means the external validity of any ranking from this benchmark is weak. Please state this limitation explicitly in Section 5 and temper the conclusion accordingly.","section":"Section 5 and Table 7"}],"minor_comments":[{"comment":"The sentence 'CLAHE is a typical baseline in this category' appears twice; please delete one occurrence.","section":"Section 3.1.1"},{"comment":"Section 1 contains an unfinished sentence, 'underwater exploration and analysis remain hampered. .', and Section 2.4 has the typo 'serval works' for 'several works'.","section":"Sections 1 and 2.4"},{"comment":"The attribution of the UIQM metric to Wang et al. (2021a) is inaccurate; UIQM was introduced by Panetta et al. and should be cited to its original source.","section":"Section 3.4"},{"comment":"Table 3 uses reference placeholders such as 'Multi-Exposure Fusion (?)' and 'Hybrid Dehazing + White Balance (?)'; these should be replaced with actual citations.","section":"Table 3"},{"comment":"In the caption, 'UCDP+HE' should read 'UDCP+HE' for consistency with the main text.","section":"Figure 10 caption"},{"comment":"Terminology is inconsistent: 'InstanceNGP' and 'Instant-NGP' are both used, and Figure 18 refers to 'UW-3DGS' while the text and Figure 19 use 'UW-GS'; please unify the names.","section":"Sections 5.1-5.3"},{"comment":"The note that the images in the first row of Figure 19 'have been enhanced for better visibility' is ambiguous; please clarify whether this enhancement is part of the compared pipeline or only for presentation, since it directly affects the fair comparison of the shown methods.","section":"Section 5.3, Figure 19"}],"recommendation":"major_revision","confidential_remarks":"The benchmark in Section 5 is presented as a contribution but is not independent of the authors' own methods: Huang et al. (2025) is unpublished and UW-GS is from the same group. I recommend that the editor require either independent replication or a clear demotion of the benchmark to a qualitative demonstration. The survey coverage itself is generally sound and could be a useful reference after revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The survey content is the real value here. The authors have organized a sprawling field—physics-based models, classical priors, CNN/GAN/Transformer/Mamba/diffusion enhancement, and photogrammetry/NeRF/3DGS reconstruction—into a coherent taxonomy with reasonably current coverage. The physics primer on the Jaffe–McGlamery model and its simplifications is accurate and accessible. For a practitioner or newcomer to underwater vision, this is a useful map.\n\nThe soft spot is Section 5, and it is not minor. The abstract promises \"quantitative and qualitative evaluations,\" but the section reports no quantitative metrics at all: no PSNR, SSIM, LPIPS, error bars, seeds, or ablations. Figure 19 compares four methods on two scenes, and the conclusion that UW-GS \"appears to be the best\" is purely visual. That would be a problem even with independent methods; it becomes a bigger problem when the best-performing method is the authors' own (UW-GS, Wang et al. 2025) and the two-stage pipeline uses their unpublished enhancement method (Huang et al., 2025). An independent reader cannot reproduce the comparison. The paper itself concedes public underwater 3D datasets are rare, but that does not reconcile the claim with the evidence. If the benchmark is illustrative, the abstract should say so and drop the word \"quantitative.\" If it is meant as a contribution, it needs actual numbers, more scenes, and released code or data.\n\nThere are also smaller issues: Table 3 contains placeholder citations (question marks) in what looks like a finished table, and there are typos like \"InstanceNGP.\" These are easy fixes, but they suggest the manuscript was not carefully proofread.\n\nThe survey chapters themselves are credible and well-cited, so I would not desk-reject the paper. Who is this for? Someone looking for a structured overview of underwater enhancement and reconstruction, or a starting point for a research project. For that purpose, it works. But the benchmark section needs to be either substantially expanded or honestly reframed before I would trust the empirical conclusions.\n\nFor peer review: yes, send it out, but tell the reviewers to focus on Section 5 and the abstract's claims. The survey deserves to exist; the benchmark claims need to be brought in line with what was actually done.","headline":"A genuinely useful survey of underwater enhancement and 3D reconstruction, but the benchmark section overclaims: the abstract promises quantitative evaluation and Section 5 delivers only visual inspection on two scenes, using the authors' own methods.","tokens_in":45647,"tokens_out":1543,"would_cite":false,"duration_ms":17508,"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":"This review makes the case that underwater visual enhancement and underwater 3D reconstruction are one coupled problem, and reports that integrated physics-based methods such as UW-GS produce the clearest novel views in the tested scenes.","keywords":["underwater image enhancement","underwater 3D reconstruction","image formation model","neural radiance fields","3D Gaussian splatting","scattering media","novel view synthesis","underwater benchmark datasets"],"falsifier":"Run a controlled comparison of UW-GS against a two-stage enhancement-plus-reconstruction pipeline on a larger, more diverse set of underwater scenes using quantitative metrics such as PSNR, SSIM, and LPIPS; if the two-stage pipeline matches or beats UW-GS on average, the paper's central ranking claim fails.","tokens_in":44741,"feed_emoji":"🌊","tokens_out":8183,"duration_ms":76615,"temperature":0.7,"pith_summary":"This paper sets out to fill a gap it identifies in the literature: no previous review covers underwater visual enhancement and underwater 3D reconstruction together. It organizes the field into a unified taxonomy, spanning physics-based models, classical image processing, deep learning, and the newer neural rendering families, and then benchmarks three reconstruction routes on public underwater datasets: reconstruction without enhancement, a two-stage enhance-then-reconstruct pipeline, and integrated models that embed water physics into reconstruction. Its central empirical claim is that the integrated route, exemplified by UW-GS, produces the clearest novel views in the tested scenes, because modeling scattering and attenuation inside the renderer beats correcting images beforehand. A sympathetic reader would take this as evidence that enhancement and geometry estimation should be treated as a single problem rather than separate steps.","feed_headline":"Integrated physics-based models win underwater 3D clarity test","feed_subtitle":"Unified review finds that fusing enhancement into reconstruction, as UW-GS does, yields the clearest novel views.","key_machinery":"The load-bearing object is the Jaffe–McGlamery underwater image formation model, which writes the observed image as the sum of direct transmission, forward scattering, and backscattering: $I(x)=I_d(x)+I_f(x)+I_b(x)$, usually simplified to $I(x)=J(x)T(x)+A(1-T(x))$. This model connects the two halves of the review: enhancement methods are categorized by how they estimate the transmission map $T$ and ambient light $A$, while reconstruction methods are categorized by whether they treat water as a medium inside the renderer or ignore it. The integrated pipelines highlighted in the benchmark, especially UW-GS, embed this model into a 3D Gaussian Splatting renderer so that scattering and attenuation are estimated per scene rather than removed in a preprocessing step.","core_discovery":"On its own terms, the paper claims to be the first systematic review to span both sides of the underwater vision problem: restoring what cameras see and reconstructing the 3D scene behind those images. It builds a taxonomy that ranges from histogram and Retinex methods through dark-channel priors and data-driven CNNs, transformers, Mamba, and diffusion models on the enhancement side, and from photogrammetry and visual SLAM through NeRF and 3D Gaussian Splatting on the reconstruction side. Its empirical section compares three pipelines: raw reconstruction, enhancement followed by reconstruction, and integrated physics-based reconstruction, using public datasets including NUSR, SeaThru, S-UW, UWBundle, and BVI-Coral. The reported outcome is that integrated models such as UW-GS render the sharpest, most color-correct novel views, that 3DGS captures fine texture better than NeRF where texture exists but blurs information-poor areas, and that dynamic NeRF variants struggle with high-frequency underwater detail. The paper frames this as evidence that the field is converging on embedding the physics of underwater light into the reconstruction itself.","pith_inferences":["Beyond the paper: if integrated medium-aware rendering keeps winning, the default for underwater novel-view synthesis is likely to shift away from two-stage enhance-then-reconstruct pipelines, making enhancement a component of the renderer rather than a separate artifact.","Beyond the paper: the reliance on a handful of small public datasets means the reported ranking is fragile; a larger multi-condition benchmark could plausibly overturn UW-GS's top placement.","Beyond the paper: the same physics-based medium modeling could be tested on downstream tasks such as underwater depth estimation, object detection, or ROV navigation, where enhanced images may or may not help depending on the task.","A testable extension is synthetic data: render scenes under several known water types and turbidities with ground-truth geometry, then compare integrated versus two-stage reconstruction quantitatively under controlled conditions."],"forward_implications":["If the paper's conclusion holds, future underwater vision systems should couple enhancement with reconstruction rather than treating image restoration as an optional preprocessing step.","Reviews and taxonomies of underwater imaging should include both enhancement and 3D reconstruction in one framework, since the physical model that explains color loss also explains reconstruction failure.","The benchmark suggests that 3D Gaussian Splatting is the more practical base for static underwater scenes, with NeRF remaining preferable for dynamic scenes.","Public underwater 3D datasets are too scarce and small to support strong generalizations, so the field's next bottleneck is data collection, not algorithms.","New methods that compare against integrated physics-based models like UW-GS will need to report whether enhancement is embedded or done in advance, because that choice affects the outcome."],"supporting_citations":[{"why":"Supplies the standard underwater image formation model that decomposes observed light into direct, forward-scattered, and backscattered components.","marker":"(Jaffe, 1990)"},{"why":"Provides the companion computer model of underwater camera systems that the simplified image formation equations are built on.","marker":"(McGlamery, 1980)"},{"why":"Revises the underwater image formation model to be physically correct; SeaThru-NeRF and integrated enhancement-reconstruction pipelines build on it.","marker":"(Akkaynak and Treibitz, 2018)"},{"why":"Introduces NeRF, the implicit neural representation that the review uses as one of the two modern reconstruction families.","marker":"(Mildenhall et al., 2020)"},{"why":"Introduces 3D Gaussian Splatting, the explicit point-based rendering method that the benchmark compares against NeRF and integrated variants.","marker":"(Kerbl et al., 2023)"},{"why":"Introduces SeaThru-NeRF, the scattering-aware NeRF baseline that supplies benchmark scenes and the physics-based integration idea.","marker":"(Levy et al., 2023)"},{"why":"Introduces UW-GS, the distractor-aware 3D Gaussian Splatting method with physics-based density control that the paper's benchmark ranks best.","marker":"(Wang et al., 2025)"},{"why":"Provides the NUSR dataset with motion masks and camera poses used to evaluate reconstruction under dynamic underwater conditions.","marker":"(Tang et al., 2024)"}],"fun_headline_variants":["Physics-based fusion beats separate steps in underwater 3D","Review: Merging enhancement with reconstruction wins underwater views","Underwater 3D: Integrated models outshine augmentation pipelines","UW-GS leads in underwater 3D clarity, systematic review finds","First unified review: Physics-aware models sharpen underwater 3D"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The benchmark conclusions assume the handful of public underwater 3D datasets used, including NUSR, SeaThru, S-UW, UWBundle, and BVI-Coral, represent the range of real underwater conditions, and that visual inspection of one scene pair is enough to rank methods like UW-GS.","fun_headline_variants_meta":{"raw":{"variants":["Physics-based fusion beats separate steps in underwater 3D","Review: Merging enhancement with reconstruction wins underwater views","Underwater 3D: Integrated models outshine augmentation pipelines","UW-GS leads in underwater 3D clarity, systematic review finds","First unified review: Physics-aware models sharpen underwater 3D"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000385,"raw_usage":{"total_tokens":2032,"prompt_tokens":935,"completion_tokens":1097,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":551,"completion_tokens_details":{"reasoning_tokens":1012}},"tokens_in":551,"tokens_out":1097,"duration_ms":7893,"temperature":1.0,"reasoning_tokens":1012,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:07:05.475212+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a controlled comparison of UW-GS against a two-stage enhancement-plus-reconstruction pipeline on a larger, more diverse set of underwater scenes using quantitative metrics such as PSNR, SSIM, and LPIPS; if the two-stage pipeline matches or beats UW-GS on average, the paper's central ranking claim fails.","supporting_citations":[],"review_version":1}