{"id":"1c3e348e-e464-448f-98f4-9dc6904c9a05","arxiv_id":"2412.15890","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"NeuroPump integrates Snell's law refraction and scattering/absorption modeling into NeRF to simultaneously correct geometry and color of underwater images, and introduces a real paired 360-degree benchmark dataset.","lead":"NeuroPump is a self-supervised method that simultaneously fixes the warped geometry and color cast in underwater photos by modeling refraction and light scattering inside a neural radiance field. It also provides a new real-world benchmark with paired underwater and in-air images for evaluating such methods. A reader interested in underwater robotics or VR might read it to see a single pipeline for both corrections and novel view synthesis.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The geometric rectification claim rests on the untested assumption s≈0 for the GoPro Hero 8; if the flat-port offset is non-negligible, the pre-rectification used for COLMAP is depth-dependent and the estimated poses become biased.","rationale":"The reader's weakest-assumption diagnosis matches the most load-bearing point in the paper. The method's novelty is explicitly the simultaneous geometric and color rectification, and the geometric rectification depends on the pre-rectification step being correct. The pre-rectification is also the only place where a wrong assumption can silently corrupt camera poses before any learning begins; the color model, by contrast, is directly validated by ablations. The concern is addressable with an empirical calibration, which is why the conditional status is appropriate. The sign error in Eq. 20 and the absence of released code/data are real reproducibility issues, but they are secondary to the geometric assumption because they do not directly threaten the internal logic of the claim. I therefore keep the reader's verdict unchanged and propose a single decisive measurement-and-rerun test.","tokens_in":17866,"tokens_out":6831,"duration_ms":68721,"concrete_test":"Measure s for the exact GoPro Hero 8 housing used in the benchmark: capture a checkerboard at three or more known depths z inside the water tank, fit the full magnification relation (Supp. Eq. 25) with s and the water refractive index as free parameters, and test whether the fitted s is statistically distinguishable from zero. If the fitted s is effectively zero (e.g., <1 mm), the concern is resolved. If s is non-negligible, re-run the entire pipeline using the depth-dependent h (Supp. Eq. 27) or a joint calibration of s with COLMAP, and compare the final PSNR/SSIM/RMSE on the published test views; if these metrics shift by more than the ablation deltas shown in Tab. 4, the central geometric claim is conditional on the housing geometry.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of simultaneous geometric and color rectification is carried by the geometric branch, which has two coupled steps: a closed-form pre-rectification of training images for COLMAP (Main Eq. 21/25, Supp. Eq. 26) and a NeRF ray-casting model that applies Snell's law. The pre-rectification factor reduces to h = tan(ϕw)/tan(ϕa) only when the perpendicular distance s between the optical center and the flat lens-case interface is assumed to be zero. The authors explicitly list this assumption as a limitation in Sec. 7, but they provide no measurement, estimate, or sensitivity analysis of s for the GoPro Hero 8 used in all experiments. The exact factor (Supp. Eq. 25) depends on both the pixel angle and the object depth z through the term (z−s); the s≈0 form removes that depth dependence. In their lab scenes, objects are at least tens of centimeters from the housing, so even a 1–2 cm port offset changes the required scale by a few percent at near depths and varies with depth. Since COLMAP pose estimation is conditioned entirely on these pre-rectified images, a systematic scale error propagates into all estimated camera poses and therefore into the jointly optimized density/color field and the final in-air rendering. Notably, the geometry-only results in Tab. 3 are only on par with the Lavest approximation, so the claimed joint advantage rests heavily on the color model; a corrupted geometry channel would undermine the 'simultaneous geometric and color rectification' claim rather than only the color component.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes NeuroPump, a self-supervised NeRF-based method that simultaneously rectifies geometric distortion (refraction at a flat lens-case interface) and color distortion (absorption and scattering) in underwater images. The model extends mip-NeRF 360 with an explicit Snell's-law ray-bending step and an underwater image formation model with per-scene attenuation coefficients and global background light, trained by a reconstruction loss plus auxiliary losses on color cast and ray distributions. The authors also introduce a real captured 360-degree benchmark with paired underwater and in-air ground-truth images across five setups. Experiments report that NeuroPump outperforms separate color and geometry rectification baselines on the joint rectification task (PSNR 22.76 vs. 20.88 for the best physics-based baseline), while geometry-only rectification is roughly on par with the Lavest radial-distortion approximation. The paper additionally demonstrates novel-view and novel-optical-parameter synthesis.","tokens_in":18227,"tokens_out":6588,"duration_ms":60844,"significance":"If the central claim holds, NeuroPump is a useful step beyond prior work that handles underwater color and geometry separately: it provides a single self-supervised pipeline that recovers an in-air scene representation from underwater multi-view images, and it releases a real paired benchmark that avoids purely synthetic evaluation. The evaluation is not circular: in-air ground truth is used only for assessment, not for fitting, since the in-air output is obtained by inverting the learned underwater formation model. The derivations in Sec. 3 and the supplementary are mostly standard and clearly presented, and the decoupling of refractive index, attenuation, and background light enables a genuinely new synthesis capability. The main weaknesses are in the geometric branch: the load-bearing s≈0 assumption for the GoPro Hero 8 is not validated, and there are algebraic inconsistencies in the pre-rectification formula and in the brightness compensation equation. These issues are fixable and do not undermine the overall idea, but they must be addressed before the paper's core claims can be fully accepted.","major_comments":[{"comment":"The geometric rectification used for COLMAP pose estimation rests on the assumption s≈0 for the GoPro Hero 8, but the paper provides no measurement of s, no upper bound, and no sensitivity analysis. In the exact factor (Supp. Eq. 25), h depends on the object depth z through (z−s); the simplified form h=tan(ϕw)/tan(ϕa) removes that depth dependence. Since COLMAP receives only pre-rectified images, a systematic warp error at near depths would bias all estimated poses and propagate into the density/color field and the final in-air rendering. The paper acknowledges the assumption as a limitation in Sec. 7, but it is load-bearing for the central 'simultaneous geometric and color rectification' claim. Please report a measured value or upper bound for s for the actual camera, and provide a sensitivity study (e.g., perturbing s by plausible amounts and reporting pose/rectification changes), or replace the assumption by joint estimation of s with poses.","section":"§4.1, Supp. C (Eq. 21/25) and Sec. 7"},{"comment":"The simplified rectification factor is inverted between the main text and the supplementary material. Main-text Eq. (21) states that for s≈0, h=tan(ϕa)/tan(ϕw), while Supp. Eq. (26) states h=tan(ϕw)/tan(ϕa). Substituting s=0 into Eq. (21) gives h=tan(ϕw)/tan(ϕa), agreeing with the supplementary derivation. The main-text expression is therefore inverted; if implemented, it would warp the pre-rectified images in the wrong direction, affecting COLMAP keypoint matching and pose estimation. Please correct the main-text formula and verify that the experiments used the correct form.","section":"§4.1 vs. Supp. C (Eq. 21 vs. Eq. 26)"},{"comment":"The brightness compensation update is written as ˆJ(ra) ← −max(1,W) ˆJ(ra). Taken literally, this multiplies the in-air radiance by a negative scale, producing an invalid inverted image rather than the brighter output that the ablation and Tab. 4 describe. The surrounding text indicates the intended operation is a positive multiplication by max(1,W) or W. Please correct the equation and clarify whether the arrow denotes an assignment in an algorithmic step, since this term is load-bearing for the color-rectification results.","section":"§6.1, Eq. (20)"},{"comment":"The joint-rectification advantage over the best separate pipeline is driven almost entirely by color: the geometry-only PSNR for Ours and Mip360(Lav) is effectively identical (21.09 vs. 21.09), with similar SSIM/RMSE. The paper should state this explicitly and provide additional geometric evidence beyond image-domain metrics, such as per-setup pose accuracy or depth error relative to the paired benchmark, to support the claim that the Snell's-law geometry branch is genuinely more accurate than the Lavest radial-distortion approximation. As written, the simultaneous-rectification claim is supported mainly by the color channel, and the geometry branch also inherits the unvalidated s≈0 assumption from the first comment.","section":"Tab. 3 and Sec. 6"}],"minor_comments":[{"comment":"The cross-references are inconsistent: the main text says 'apply Eq. 25' and the supplementary refers to 'main paper's Eq. 25', but the equations are numbered Eq. (21) in the main text and Eq. (26) in the supplementary. Please harmonize the equation numbering throughout.","section":"§4.1 and Supp. C"},{"comment":"The use of '←' inside a displayed mathematical formula is unusual; if this denotes an algorithmic assignment, please use explicit pseudocode or a clear notation to avoid ambiguity with mathematical equality.","section":"§6.1, Eq. (20)"},{"comment":"The sentence 'We present a the first model that simultaneously restores both geometric and color distortions' contains a typo ('a the'); please correct it.","section":"§7"},{"comment":"The note that the identical value of Ours and Mip360 (Lav) + Others baselines is a coincidence is confusing as printed; please rephrase or remove it, since the table caption already explains the grouping.","section":"Tab. 3"},{"comment":"The phrase 'to our best knowledge' should be 'to the best of our knowledge' for grammatical correctness.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The algebraic inversion in Eq. (21)/Supp. (26) and the negative sign in Eq. (20) are concerning because they are central to the method, but they look like typographical errors rather than fundamental flaws; the authors should confirm against their released code which form was actually used. The s≈0 assumption is the most substantively important issue: it is load-bearing for the geometric branch and is only acknowledged, not validated. If the authors can supply a measurement or a convincing sensitivity analysis, and correct the equations, the paper would be a solid contribution given the useful real paired benchmark and the self-supervised formulation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: this is a solid paper that deserves a serious referee. The authors build the first NeRF-style pipeline that jointly rectifies underwater geometry (refraction) and color (scattering/absorption), and they back it with a new real paired underwater 360-degree dataset. The central claim—that joint rectification beats doing geometry and color separately—is supported by their experiments: on the final in-air target, Ours reaches PSNR 22.76 versus 20.88 for the strongest physics-based baseline, and the ablations show both brightness compensation and the color-cast loss help. That is a real, if modest, improvement.\n\nWhat is actually new: integrating Snell's law refraction into the NeRF ray-casting and camera-pose pre-rectification, plus the decoupled water parameters (β, A) that enable novel-view and novel-optical-parameter synthesis. The real paired dataset is the most immediately useful artifact; no prior benchmark has in-air ground truth aligned with underwater 360 views. The derivations in the main text and supplementary are mostly standard and clear.\n\nThe soft spots are real but not fatal. The biggest is the s≈0 assumption for the GoPro Hero 8's flat lens case. The stress-test note is correct that the full pre-rectification factor depends on depth when s is non-negligible, and COLMAP poses are estimated from images pre-rectified with the s≈0 form. The authors explicitly acknowledge this in Section 7 but give no measurement of s for their camera and no sensitivity analysis. That said, the geometry-only results are on par with the Lavest radial-distortion approximation, which suggests the practical effect in their lab setup is small; the joint advantage comes mostly from the color model. Still, the 'simultaneous geometric rectification' claim should be hedged until s is measured or the sensitivity is characterized.\n\nThere is also a sign error in Eq. 20: the brightness compensation is written as a negative multiplication, which would flip the image. It is clearly a typo (the ablation shows it helps), but it will confuse any reader reproducing the method. Code and data are not available from the manuscript, and there are no error bars, so the quantitative results are not fully reproducible as presented.\n\nRecommendation: send to peer review. The method is original in its integration, the dataset is valuable, and the flaws are addressable. Require the authors to measure or bound s for their camera, release code/data, fix the Eq. 20 typo, and add per-scene variance or a small sensitivity analysis. With those, this would be a strong contribution to the underwater vision and NeRF communities.","headline":"A genuinely useful joint underwater rectification pipeline with a real paired dataset; the geometry branch rests on an acknowledged but unquantified s≈0 assumption, so treat the geometric claims as conditional pending sensitivity analysis.","tokens_in":734,"tokens_out":2006,"would_cite":true,"duration_ms":36539,"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 paper claims that underwater photographs can have their refraction-warped geometry and water-discolored colors corrected in a single self-supervised neural rendering pass.","keywords":["underwater image restoration","geometric distortion rectification","color distortion rectification","neural radiance fields","Snell's law refraction","underwater benchmark dataset","novel view synthesis"],"falsifier":"Take a flat-port underwater camera whose $s$ has been measured, capture a scene with known 3D structure, and compare poses and depths obtained with the simplified factor $h = \\tan(\\phi_w)/\\tan(\\phi_a)$ against those obtained with the full expression; a systematic divergence without another error source would refute the $s \\approx 0$ basis of the geometric claim.","tokens_in":17673,"feed_emoji":"🌊","tokens_out":9904,"duration_ms":80675,"temperature":0.7,"pith_summary":"NeuroPump is a self-supervised pipeline that takes multiple underwater views of one scene and recovers the scene as it would appear in air: straight geometry and true colors, in a single optimization rather than two separate fixes. The authors claim this is the first method to correct underwater geometric distortion and color distortion simultaneously. The idea is to build refraction (Snell's law) and water absorption/scattering directly into a neural radiance field, so the model learns an in-air 3D representation while also estimating water parameters. To evaluate it, the paper contributes a 360-degree benchmark with real paired underwater and in-air images across five scenes. If the claim holds, underwater 3D capture with an ordinary flat-port camera no longer needs a dome port, and the same trained scene can be re-rendered with different water conditions.","feed_headline":"One model fixes underwater geometry and color together","feed_subtitle":"It learns refraction and water attenuation inside a neural radiance field, then renders the scene without them.","key_machinery":"The load-bearing mechanism is a volume-rendering model that treats water as two decoupled physical effects. Each camera ray is bent once at the flat lens-case interface by Snell's law, giving the refracted ray $r_w$; the accumulated color follows the simplified scattering-plus-absorption model $I(r_w) \\approx e^{-\\beta d}J(r_w) + (1-e^{-\\beta d})A$, where $\\beta$ is the per-channel water attenuation, $A$ is the global background light, $d$ is the distance to the first opaque surface, and $J$ is the unattenuated in-air radiance. The network learns the radiance field $(c_o, \\sigma_o)$ together with $\\beta$ and $A$; rendering without the water terms yields the rectified in-air image, while the simplified pre-rectification factor $h = \\tan(\\phi_w)/\\tan(\\phi_a)$ (valid when the optical-center-to-interface distance $s \\approx 0$) rewrites underwater images before pose estimation.","core_discovery":"The paper's central claim is that a single neural renderer can recover both the geometric and the color corrections for underwater images at once. Instead of removing refraction with a dome port or approximating it as radial distortion, NeuroPump bends each camera ray once at the flat lens-case interface using Snell's law, then renders color with a simplified underwater formation model $I \\approx e^{-\\beta d}J + (1-e^{-\\beta d})A$, where $\\beta$ is per-channel water attenuation, $A$ is global background light, and $J$ is the unattenuated in-air radiance. After training, turning off the water terms yields a rectified in-air image and depth, and changing the decoupled optical parameters yields synthetic novel views with different water properties. The paper reports that on its new real paired 360-degree benchmark, this joint rectification is quantitatively and qualitatively closer to the captured in-air ground truth than sequential pipelines that first fix refraction and then apply a color-restoration method.","pith_inferences":["A natural next test is to estimate the optical-center-to-interface distance $s$ instead of assuming it is zero; the paper itself lists joint estimation of intrinsics, extrinsics, and water parameters as future work.","Because the water parameters are learned separately from the scene radiance, a model trained in clear lab water could in principle be re-rendered with field-measured attenuation and background light, moving the approach from tank to open water.","The limit imposed by feature-matching pose estimation suggests that turbid-water failure is as much a pose-estimation problem as a rendering problem; a pose-free variant would isolate how much of the geometric gain comes from Snell's law itself."],"forward_implications":["Underwater 3D capture with a plain flat-port camera can recover geometrically correct depth and a color-true in-air appearance without a dome port.","A single trained scene representation supports novel-view synthesis and lets a user change refractive index, background light, and attenuation to simulate different water conditions.","Joint rectification, rather than separate color and geometry pipelines, becomes a viable default for underwater neural radiance field reconstruction.","The new real paired 360-degree benchmark gives future methods a way to measure both geometry and color rectification against true in-air images."],"supporting_citations":[{"why":"It supplies the base anti-aliased neural radiance field architecture and the Lprop/Ldist losses on which NeuroPump builds.","marker":"[6]"},{"why":"It provides Snell's law, used to bend rays at the lens-case interface.","marker":"[24]"},{"why":"It gives the classic scattering and absorption image-formation model simplified into Equation (11).","marker":"[26]"},{"why":"It supplies the global brightness-compensation step that restores unattenuated in-air radiance.","marker":"[2]"},{"why":"It is the refraction-removal baseline, using a radial-distortion approximation, against which NeuroPump compares poses.","marker":"[34]"},{"why":"It is the neural-radiance-field-in-scattering-media baseline whose reconstruction loss NeuroPump adapts.","marker":"[36]"},{"why":"It is the structure-from-motion pose estimator whose refraction-sensitive input motivates the pre-rectification step.","marker":"[52, 53]"}],"fun_headline_variants":["Underwater photos get geometry and color fixed at once","NeuroPump: one network rectifies underwater geometry and color","Self-supervised NeRF pumps out water distortion in one pass","Joint underwater correction: geometry and color without separate steps"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The geometric correction rests on treating the perpendicular distance between the camera's optical center and the flat lens-case window as negligible ($s \\approx 0$); if that distance is not negligible for a given camera, the pre-rectified images fed to pose estimation are warped by the wrong factor and the recovered geometry is biased.","fun_headline_variants_meta":{"raw":{"variants":["Underwater photos get geometry and color fixed at once","NeuroPump: one network rectifies underwater geometry and color","Self-supervised NeRF pumps out water distortion in one pass","Joint underwater correction: geometry and color without separate steps"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00025,"raw_usage":{"total_tokens":1543,"prompt_tokens":926,"completion_tokens":617,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":542,"completion_tokens_details":{"reasoning_tokens":550}},"tokens_in":542,"tokens_out":617,"duration_ms":5880,"temperature":1.0,"reasoning_tokens":550,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T10:59:46.823108+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a flat-port underwater camera whose $s$ has been measured, capture a scene with known 3D structure, and compare poses and depths obtained with the simplified factor $h = \\tan(\\phi_w)/\\tan(\\phi_a)$ against those obtained with the full expression; a systematic divergence without another error source would refute the $s \\approx 0$ basis of the geometric claim.","supporting_citations":[{"cited_title":"Computer modeling and the design of opti- mal underwater imaging systems","cited_arxiv_id":null,"evidence_quote":"It gives the classic scattering and absorption image-formation model simplified into Equation (11)."},{"cited_title":"Mip-nerf 360: Unbounded anti-aliased neural radiance fields","cited_arxiv_id":null,"evidence_quote":"It supplies the base anti-aliased neural radiance field architecture and the Lprop/Ldist losses on which NeuroPump builds."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It provides Snell's law, used to bend rays at the lens-case interface."},{"cited_title":"Sea-thru: A method for removing water from underwater images","cited_arxiv_id":null,"evidence_quote":"It supplies the global brightness-compensation step that restores unattenuated in-air radiance."},{"cited_title":"Underwater camera calibration","cited_arxiv_id":null,"evidence_quote":"It is the refraction-removal baseline, using a radial-distortion approximation, against which NeuroPump compares poses."},{"cited_title":"Seathru- nerf: Neural radiance fields in scattering media","cited_arxiv_id":null,"evidence_quote":"It is the neural-radiance-field-in-scattering-media baseline whose reconstruction loss NeuroPump adapts."}],"review_version":1}