{"id":"aabaf0f1-7a5a-47f9-b529-9fc5e2e6629e","arxiv_id":"2504.19289","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A dataset built from real marine snow masks extracted by median frame subtraction enables a video enhancement network to reduce snow and improve SLAM feature matching and dense mapping.","lead":"Researchers generated a paired training set for removing marine snow by subtracting a time-averaged frame from real underwater video to extract snow particles, then overlaying those particles onto clean frames. They trained a video enhancement network on these pairs and report that it improves feature matching and 3D reconstruction in a proprietary SLAM system.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SLAM improvement claim rests on a single unquantified average over eight proprietary sequences; without per-sequence values or a paired test, the headline gain may be noise.","rationale":"The reader identified mask purity as the weakest assumption, and I agree that the absence of a quantitative mask-purity analysis is a real concern. However, I see a more immediately load-bearing issue in the evaluation: the pseudo-GT construction could be imperfect and still yield a useful enhancement, but the central SLAM claim cannot be accepted on a single unquantified average. Table I is the only hard number supporting the headline, and it is a mean over eight sequences with no reported spread. Figure 6 says 'consistently higher' but shows smoothed curves with no aggregate statistic. Because the data and pipeline are proprietary, independent reproduction is impossible. The concrete test above would settle whether the 28% dense-mapping improvement is robust. This supports, rather than changes, the reader's CONDITIONAL verdict.","tokens_in":5773,"tokens_out":6415,"duration_ms":74207,"concrete_test":"Ask the authors for the per-sequence dense-map point counts for all eight evaluation sequences. Compute the paired difference (ours minus no-enhancement), its 95% bootstrap confidence interval, and a paired Wilcoxon signed-rank test; repeat with each sequence removed (jackknife). If the confidence interval includes zero, the Wilcoxon p-value is not below 0.05, or any single-sequence removal reverses the sign of the mean difference, the claim of a higher average number of reconstructed points is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is empirical: processing with the pseudo-pair-trained network improves SLAM. For this to hold, the reported improvements must be real and reproducible. The least secure point is the evidence base in Section IV-C. Table I reports one mean dense-map point count per method over eight sequences (893,076 vs 1,142,521.5) with no standard deviation, no per-sequence breakdown, and no statistical test. Figure 6, used for 'consistently achieves a higher number of frame-to-frame matches,' is a smoothed curve with no aggregate statistic or error bars, and Figure 4 uses smoothing as well. With only eight sequences and a proprietary SLAM/dense-mapping pipeline, a single outlier sequence could move the average by the reported ~28% margin, and the result cannot be independently reproduced or audited. The paper itself notes in Section IV-B that standard pixel-level metrics may not reflect SLAM performance; the same caution applies to unreplicated, variance-free SLAM runs. The median-subtraction mask-purity issue is a genuine concern for generalization, but the variance-free evaluation is the direct threat to this specific claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a self-supervised data-generation pipeline for marine snow removal. Static background patches are selected from raw underwater videos, a temporal median frame is subtracted from each frame to obtain dynamic snow masks (Eq. 2), and these masks are additively overlaid onto clean reference frames to create paired snowy/snow-free training data (Eq. 3). The authors then train an enhancement network based on BVI-Mamba on these internally generated pseudo-pairs and evaluate the enhanced videos by feeding them into a proprietary SLAM and dense-mapping system. The reported claims are that the method reduces snow-like keypoints, consistently increases frame-to-frame feature matches, and raises the average dense-map point count from 893,076 to 1,142,521.5 across eight real sequences.","tokens_in":5952,"tokens_out":3046,"duration_ms":34252,"significance":"If the central claim is substantiated, the paper makes a useful practical contribution: it avoids fully synthetic snow simulation by reusing real snow artifacts, enables supervised training without manually cleaned ground truth, and evaluates the enhancement through downstream SLAM metrics rather than pixel-level metrics that the authors rightly argue are not meaningful for this task. The idea of internally generated pseudo-pairs from median subtraction is simple and potentially transferable to other transient-noise removal problems. The main value is therefore conditional on whether the reported SLAM improvement is statistically real and whether the mask-extraction assumptions hold in the deployment conditions. The paper does not provide machine-checked proofs or code, but the described pipeline is straightforward enough that the missing per-sequence statistics and mask-purity measurements could be supplied in a revision.","major_comments":[{"comment":"The headline quantitative claim rests entirely on a single average dense-map point count over eight sequences, with no standard deviation, no per-sequence values, and no paired significance test. With n = 8, one favorable sequence could move the average by roughly the reported 28% margin, so the claimed improvement from 893,076 to 1,142,521.5 points is not yet supported. Please report the per-sequence point counts, the mean and spread for each method, and a paired statistical test (e.g., Wilcoxon signed-rank or a bootstrap confidence interval on the mean difference). Without this, the central claim that enhancement improves SLAM is not independently auditable.","section":"Section IV-C, Table I"},{"comment":"The mask-extraction step assumes that the scene is static over the median window and that marine snow is the only dynamic signal. If the camera drifts, the vehicle moves, or bubbles, fish, or other objects pass through the selected patch, the subtraction in Eq. (2) leaves scene residuals in the mask, and the network is trained to erase real structure. Manual patch selection reduces this risk but does not quantify it. The paper should provide a quantitative purity analysis of the extracted masks, for example residual energy outside detected snow components, or a comparison of masks with manually annotated snow regions, and should state how sensitive the results are to the median window length N and patch size.","section":"Section III-A, Eq. (2)"},{"comment":"The claim that the method 'consistently achieves a higher number of frame-to-frame matches' is supported only by smoothed curves without aggregate statistics, per-sequence breakdowns, or error bars. Figure 6 in particular cannot substantiate 'consistently' when the underlying per-sequence values are not shown. Please provide a table of frame-to-frame match counts (or match-rate ratios) for each of the eight sequences, with means and confidence intervals, and state how the smoothing was applied and over what window. The same applies to the keypoint-count claim in Fig. 4, where 'lower is better' is asserted but no statistical summary is given.","section":"Section IV-C, Figs. 4 and 6"},{"comment":"The paper correctly argues that pixel-level metrics are not the right evaluation for this task, but it then substitutes a single proprietary SLAM/dense-mapping pipeline and a small set of eight sequences as the only evidence. Because the SLAM system and the underwater sequences are proprietary, the result cannot be independently reproduced. At minimum, the authors should report the variance across sequences and ideally release the per-sequence metrics or a runnable evaluation protocol. This is load-bearing because the entire contribution is the claimed improvement in downstream task performance.","section":"Section IV-B and Section IV-C"}],"minor_comments":[{"comment":"The phrase 'snow, free underwater videos' should be 'snow-free underwater videos', and 'with the absence of groundtruth' should be rephrased as 'in the absence of ground truth'.","section":"Abstract"},{"comment":"The phrase 'loosing details' should be 'losing details'.","section":"Section I"},{"comment":"The index notation is confusing: I_t^snow is indexed by t, while the clean frame is indexed by t' and the spatial offsets are Delta x, Delta y. Please define the ranges of t, t', and the sliding window, and clarify whether the clean frame and the mask are temporally aligned within the window or chosen independently.","section":"Section III-A, Eq. (3)"},{"comment":"The architecture is called BVI-Mamba but the cited reference [18] is titled 'BVI-RLV: A fully registered dataset and benchmarks for low-light video enhancement'. If BVI-Mamba is a distinct method or is detailed elsewhere, please cite the correct source or clarify the relationship.","section":"Section III-B and References"},{"comment":"The text describes two reconstructed sequences (2023 000015 000135 and 2022 005644 005744) but the figure caption names only 2022 004742 004844. Please align the caption with the sequences actually shown and specify which point clouds correspond to which sequence.","section":"Figure 7, Section IV-C"},{"comment":"The patch selection for static background regions is described only as 'fixed-size patches of 550 x 600 pixels' that 'avoid any foreground objects'. Please state how many patches were selected, from how many source sequences, and what manual criteria were used, so that the dataset composition is reproducible.","section":"Section IV-A"}],"recommendation":"major_revision","confidential_remarks":"The paper's practical contribution is plausible, but the proprietary data and SLAM pipeline, combined with the small n=8 evaluation and the absence of variance or significance measures, make the central claim difficult to verify. I would encourage the editor to ask for the per-sequence numbers and a mask-purity analysis; if those are provided and are favorable, the paper could become acceptable. The fit to a computer-vision venue is reasonable given the dataset-generation and task-based evaluation focus, though the manuscript would benefit from a clearer statement of limitations in the main text, beyond the brief acknowledgement in Section IV-B."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The new thing here is real: extracting marine snow masks by median frame subtraction from raw footage and overlaying them on clean frames to build training pairs. That is a clean, practical workaround for the missing paired-data problem, and it is clearly different from the synthetic snow in [10] and the hand-collected samples in [17]. The authors also deserve credit for evaluating on downstream SLAM instead of pixel metrics, and for explicitly acknowledging that their pseudo-GT is only the cleanest available data, not true ground truth. That is honest framing.\n\nThe soft spots are real but not disqualifying. The SLAM evidence is the load-bearing piece, and right now it is thin: eight proprietary sequences, a single average dense-map point count in Table I (893,076 vs 1,142,521.5) with no per-sequence breakdown, no variance, no paired test. Figure 6 is smoothed and has no aggregate statistic. With only eight sequences, one outlier could plausibly move the average by that margin, so the stress-test note lands. The mask-purity assumption is also unquantified: subtracting a median assumes a static background behind the snow, and although manual patch selection reduces the risk, the paper does not show how much scene residual remains. That matters for generalization, though it is a secondary concern relative to the missing error bars.\n\nWhat the paper does well outweighs these issues. The dataset generation method is novel, reproducible in principle, and likely useful for other underwater video tasks beyond SLAM. The enhancement architecture is off-the-shelf, but that is fine; the contribution is the training data, not the network. The citation pattern looks reasonable, with prior marine-snow work clearly distinguished.\n\nMy overall take: the central claim—that this pseudo-pair training improves SLAM—is plausible but under-supported as presented. The fix is straightforward: report per-sequence results, include standard deviations or a paired significance test, and ideally release or anonymize a subset of the data so others can verify. The mask-purity question could be addressed with a small quantitative analysis of residual background energy.\n\nWho is this for? Researchers working on underwater robotics, marine snow removal, or SLAM under degraded visual conditions. It deserves a serious referee, but it needs revision before the improvement claim is fully convincing. I would send it to review with a request for stronger statistics and more transparency about the proprietary data.","headline":"A genuinely useful pseudo-pair trick for marine snow training data, with a SLAM evaluation that is directionally right but statistically too thin to fully land the headline claim.","tokens_in":6477,"tokens_out":948,"would_cite":true,"duration_ms":11170,"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":"Real marine snow masks extracted from video can train an enhancer that improves underwater SLAM.","keywords":["marine snow removal","underwater video enhancement","pseudo ground truth","median frame subtraction","simultaneous localization and mapping","SLAM feature matching","video restoration","dense mapping"],"falsifier":"Take a snow-free static scene, introduce a small continuous camera pan during the clip, apply the paper's median-subtraction pipeline, and train the enhancement network on the resulting pairs; if the network systematically removes or blurs stationary edges rather than only particles, the mask is contaminated. A second check is to measure mask purity directly by recording the same underwater scene with two exposures, using the short exposure to identify true snow particles and computing what fraction of the extracted mask pixels overlap true particles.","tokens_in":5576,"feed_emoji":"🌊","tokens_out":4871,"duration_ms":45849,"temperature":0.7,"pith_summary":"The paper claims that marine snow in underwater video can be suppressed by training a video enhancement network on internally generated pseudo ground truth, eliminating the need for manually paired clean and snowy footage. The dataset trick is to take real snow-affected clips, compute a per-pixel median over a short frame window as an estimate of the static background, subtract that median from each frame to isolate moving snow particles, and overlay the resulting masks onto clean reference patches at random positions and times. A video enhancement network (BVI-Mamba) trained on these pairs is then applied to full frames before running simultaneous localization and mapping (SLAM) and dense mapping. On eight real underwater sequences, the paper reports fewer snow-like keypoints, more frame-to-frame feature matches, and an average dense-map point count of 1,142,521.5 versus 893,076 without enhancement, supporting the idea that whole-frame enhancement can improve geometric vision without modifying the SLAM pipeline.","feed_headline":"Real video snow trains a net that boosts underwater SLAM","feed_subtitle":"Pseudo-pairs from median subtraction raise dense-map points from 893k to 1.14M on eight real sequences.","key_machinery":"The mechanism is median-subtracted snow-mask extraction and overlay. A median frame over $N$ frames estimates the static background in a snow-affected clip; subtracting it from each frame leaves only the dynamic bright particles, which are then overlaid onto clean reference patches at randomised temporal and spatial offsets to form paired training data. The second load-bearing component is BVI-Mamba, a video enhancement architecture with feature-level frame alignment and 2D Selective Scan state-space modelling, which is trained with an $\\ell^1$ loss to map snow-affected frames to the clean references.","core_discovery":"The central claim is that real marine snow artefacts extracted by median-frame subtraction can serve as high-quality training signal, and that a network trained on these internally generated pairs generalises to unseen footage and improves feature-based underwater SLAM. The assertion is that because the training masks come from real video rather than synthetic simulation, the model learns the true appearance and motion of marine snow, and because enhancement acts on entire frames rather than by rejecting keypoints, the downstream mapping pipeline needs no modification. The evidence is a comparative evaluation on eight real sequences where the proposed enhancement yields a lower number of detected keypoints per frame, interpreted as fewer snow detections, a consistently higher number of frame-to-frame feature matches, and a denser 3D reconstruction, with the average dense-map point count rising about 28% over the no-enhancement baseline.","pith_inferences":["If the median-subtraction masks are as clean as assumed, the method turns unlabelled underwater video into effectively unlimited paired training data; a natural extension is to apply the same pipeline to other transient visual noise such as bubbles, sediment plumes, or drifting algae that move against a static background.","Part of the dense-map point increase could come from the network smoothing low-texture seabed and generating spurious matches rather than from recovering true structure; a test on a synthetic scene with known geometry and simulated snow would separate snow suppression from detail hallucination.","The approach is likely sensitive to camera motion and vehicle manoeuvring, since the static-background assumption degrades as motion grows; a controlled sweep of camera translation speed would quantify the ceiling of the method."],"forward_implications":["Preprocessing underwater inspection footage with this enhancement should increase the number of valid frame-to-frame feature matches, improving feature-tracking stability in SLAM.","Dense 3D reconstruction should become more complete, since the paper's average reconstructed point count rises from 893,076 to 1,142,521.5 with enhancement.","Because enhancement is applied to entire frames rather than by filtering keypoints, the same preprocessing can benefit any downstream underwater vision task that consumes raw imagery, such as object detection or inspection.","The dataset-generation procedure is portable: it only needs a snow-affected clip and a clean reference patch, so it can be re-run on footage from different sites, depths, or cameras to build customised paired datasets."],"supporting_citations":[{"why":"Supplies the handcrafted spatiotemporal-filtering baseline that motivates the learning-based approach.","marker":"[16]"},{"why":"Provides the fully synthetic marine snow dataset whose generalisation limits the paper argues against.","marker":"[10]"},{"why":"Represents the keypoint-rejection alternative for SLAM that the paper's whole-frame enhancement avoids.","marker":"[17]"},{"why":"Defines the BVI-Mamba architecture used as the video enhancement network.","marker":"[18]"},{"why":"Underlies the 2D Selective Scan state-space backbone used inside BVI-Mamba.","marker":"[19]"},{"why":"Supplies the Bayesian enhancement comparison baseline in the SLAM evaluation.","marker":"[15]"}],"fun_headline_variants":["Real snow pseudo-pairs from video boost underwater SLAM","Video-derived snow pairs raise SLAM map density 28%","Internally generated snow pairs improve underwater feature matching","Pseudo ground truth from marine snow aids SLAM reconstruction","Snow from raw video trains net, densifies underwater maps"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the median of a frame window isolates marine snow, which holds only if the scene behind the snow is static for the whole window and the only moving signal is snow; any camera drift, vehicle motion, or passing non-snow object contaminates the mask with real scene structure that the network is then trained to erase.","fun_headline_variants_meta":{"raw":{"variants":["Real snow pseudo-pairs from video boost underwater SLAM","Video-derived snow pairs raise SLAM map density 28%","Internally generated snow pairs improve underwater feature matching","Pseudo ground truth from marine snow aids SLAM reconstruction","Snow from raw video trains net, densifies underwater maps"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000476,"raw_usage":{"total_tokens":2303,"prompt_tokens":827,"completion_tokens":1476,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":443,"completion_tokens_details":{"reasoning_tokens":1396}},"tokens_in":443,"tokens_out":1476,"duration_ms":13539,"temperature":1.0,"reasoning_tokens":1396,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:55:40.582788+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a snow-free static scene, introduce a small continuous camera pan during the clip, apply the paper's median-subtraction pipeline, and train the enhancement network on the resulting pairs; if the network systematically removes or blurs stationary edges rather than only particles, the mask is contaminated. A second check is to measure mask purity directly by recording the same underwater scene with two exposures, using the short exposure to identify true snow particles and computing what fraction of the extracted mask pixels overlap true particles.","supporting_citations":[{"cited_title":"Real-time marine snow noise removal from underwater video sequences,","cited_arxiv_id":null,"evidence_quote":"Supplies the handcrafted spatiotemporal-filtering baseline that motivates the learning-based approach."},{"cited_title":"Marine snow simulation and elimination in video,","cited_arxiv_id":null,"evidence_quote":"Provides the fully synthetic marine snow dataset whose generalisation limits the paper argues against."},{"cited_title":"Detecting and suppressing marine snow for underwater visual slam,","cited_arxiv_id":null,"evidence_quote":"Represents the keypoint-rejection alternative for SLAM that the paper's whole-frame enhancement avoids."},{"cited_title":"Bayesian Neural Networks for One-to-Many Mapping in Image Enhancement","cited_arxiv_id":"2501.14265","evidence_quote":"Supplies the Bayesian enhancement comparison baseline in the SLAM evaluation."}],"review_version":1}