{"id":"908b022b-4213-4755-947d-fd508a946093","arxiv_id":"2509.08490","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A structured review of underwater object detection finds that synthetic DALL-E 3 images give tiny mAP gains and that a LoRA-fine-tuned Florence-2 localizes objects but hallucinates class names.","lead":"This paper reviews the challenges of underwater object detection and tests two modern ideas: generating synthetic training images with DALL-E 3 and fine-tuning the Florence-2 vision-language model with LoRA. Real gains are small, and the Florence-2 experiment could not be scored because the model produced misspelled class names.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 8's synthetic-data gain is within single-run noise and is confounded with training-set size; the reported metrics cannot support the claim that DALL-E 3 augmentation helps.","rationale":"The reader's verdict of CONDITIONAL is appropriate, and the weakest assumption identified by the reader matches the central problem I found: the Table 8 comparison lacks statistical support and is confounded by dataset size. My stress-test did not reveal an additional fatal flaw in the survey portion of the paper; the review's organizational content is largely a recapitulation of prior work, and the main risk is in the two original case studies. For the YOLO11 experiment, the reported deltas (mAP@50 +0.003, mAP@50-95 +0.004) are far smaller than typical run-to-run variation for detection models, and without seeds or variance there is no basis to distinguish signal from noise. For the Florence-2 experiment, the paper itself reports that quantitative evaluation was impossible due to hallucinated class names, so the 'strong localization' claim rests on selected qualitative examples. This concern is internal to the paper's argument, not a disagreement with external consensus. The proposed concrete test is a controlled ablation that would settle whether the synthetic-data augmentation effect is real; until that is done, the manuscript should be accepted only conditionally on reframing the case-study claims as anecdotal and providing the missing statistical support or artifacts.","tokens_in":34360,"tokens_out":2688,"duration_ms":351070,"concrete_test":"Retrain YOLO11 on three conditions with identical hyperparameters, optimizer, and evaluation split: (A) the original 7,600 real images, (B) 8,800 real images (e.g., by adding 1,200 additional real underwater images from the same pool or by duplicating/augmenting real images to match the size increase), and (C) the 8,800-image combined real+synthetic dataset. Run each condition with at least 5 random seeds and report mean +/- standard deviation for mAP@50, mAP@50-95, recall, and precision. If condition C is not significantly better than condition B (e.g., by a paired bootstrap or Wilcoxon test), then the improvement in Table 8 cannot be attributed to synthetic data, and the claim should be relabeled as an uncontrolled pilot observation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central load-bearing empirical claim is in Section 5.1.3, Table 8: training YOLO11 on 8,800 combined images (7,600 real + 1,200 DALL-E 3 synthetic) yields mAP@50 0.796 vs 0.793, mAP@50-95 0.505 vs 0.501, recall 0.736 vs 0.714, and precision 0.780 vs 0.805 relative to training on 7,600 real images. The paper reports no number of seeds, no variance, no confidence intervals, and no statistical test. More importantly, the comparison is confounded: the combined condition changes both the data source and the dataset size (7,600 -> 8,800), so any observed improvement could come from simply having more training images or from altered class balance, not from the synthetic images specifically. The paper's own discussion in Section 5.1.4 acknowledges that the performance gap is small and that 1,200 synthetic images 'might not have been sufficient,' which undercuts the earlier statement that the results 'highlight the strength of synthetic data augmentation.' The Florence-2 case study in Section 5.2.3 similarly claims 'strong localization capabilities' and that the model 'excelled' at drawing bounding boxes, but the paper itself states that hallucinated class names made it impossible to compute meaningful mAP or recall; without a quantitative or carefully controlled qualitative evaluation, that claim is anecdotal. Because the two case studies are presented as part of the review's original contribution, this unsupported quantitative claim is the most load-bearing weakness.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript is a structured literature review of underwater object detection (UOD). It categorizes challenges into five areas—image quality degradation, target-related issues, data-related challenges, computational and processing constraints, and detection methodology limitations—and surveys solutions from traditional image enhancement and restoration through modern CNN-, transformer-, and hybrid-based detectors, culminating in a discussion of large vision-language models (LVLMs). The review contributes two original case studies: (1) augmenting a real underwater dataset (7,600 images) with 1,200 DALL-E 3 synthetic images and evaluating YOLO11 on the combined 8,800-image set, and (2) fine-tuning Florence-2 with LoRA on the same combined data. The paper claims that synthetic augmentation modestly improves YOLO11 detection metrics and that the LoRA-fine-tuned Florence-2 demonstrates strong localization, while also openly acknowledging limitations including the small performance gap and hallucinated class names.","tokens_in":34805,"tokens_out":3341,"duration_ms":31693,"significance":"If the empirical claims hold, the paper would provide a useful reference point for applying generative LVLMs and parameter-efficient fine-tuning to data-scarce UOD. The literature survey is broad and the taxonomy is sensible, and the authors are commendably explicit about the limitations of their own case studies. The main significance is therefore as a structured survey with two preliminary feasibility pilot studies; the quantitative conclusions of those pilots are not yet established at the level the paper asserts.","major_comments":[{"comment":"The central claim that adding 1,200 DALL-E 3 synthetic images improves YOLO11 detection is not supported by the reported numbers. The differences in mAP@50 (0.796 vs. 0.793) and mAP@50-95 (0.505 vs. 0.501) are within the typical single-run stochastic variation of YOLO training, while recall (0.736 vs. 0.714) and precision (0.780 vs. 0.805) move in opposite directions. No training seeds, variance estimates, confidence intervals, or significance tests are reported. In addition, the comparison is confounded: the combined condition changes both the data source and the dataset size (7,600 to 8,800 images), so any observed improvement could result from simply having more training images or altered class composition rather than from synthetic data specifically. To support the stated conclusion, the authors should report multiple seeds with mean and standard deviation and include an ablation that isolates the synthetic contribution at a matched dataset size (e.g., adding 1,200 real images or subsampling the combined set to the original size).","section":"Section 5.1.3, Table 8"},{"comment":"The discussion explicitly concedes that 'the limited number of synthetic images (1,200) might not have been sufficient to blend effectively with the larger real dataset (7,600).' This admission directly undercuts the stronger statement in Section 5.1.3 that the results 'highlight the strength of synthetic data augmentation.' As written, the paper is internally inconsistent about the strength of its own evidence. The conclusion should either be tempered to reflect that no reliable improvement was demonstrated, or the experimental evidence must be strengthened to justify the stronger claim.","section":"Section 5.1.4"},{"comment":"The claim that the LoRA-fine-tuned Florence-2 'demonstrated strong localization capabilities' and 'excelled' at drawing bounding boxes is not supported by any quantitative evaluation. The paper itself states that hallucinated class names made it impossible to compute meaningful mAP or recall values. Without either a controlled qualitative protocol (e.g., a predefined set of images, independent human rating, or localization-only IoU computed independently of class labels), this positive characterization remains anecdotal. The authors should either provide such an evaluation or present the case study strictly as a feasibility demonstration with no claims about the strength of localization.","section":"Section 5.2.3"},{"comment":"The paper claims that the quality of the enhanced synthetic images was 'rigorously evaluated using metrics such as PSNR and SSIM,' but no PSNR or SSIM values are reported anywhere in the manuscript. This makes the quality claim unverifiable. The authors should report the actual metric values and describe the reference images and protocol used, or remove the claim.","section":"Section 5.1.4"}],"minor_comments":[{"comment":"The pipeline defines three paths—P1 (synthetic only), P2 (real only), and P3 (mixed)—but Section 5.1.3 only evaluates P2 versus P3. P1 is never used in the experiments; either evaluate the synthetic-only condition or remove it from the pipeline description to avoid confusion.","section":"Section 5.1, Fig. 6"},{"comment":"The abbreviation for Detection Transformer is written as 'DETR' in most places but as 'DeTR' in at least one instance and in the Fig. 5 caption. Please standardize the spelling.","section":"Section 3.3.1 and Fig. 5"},{"comment":"Equation (A.3) reads 'V = 255 · G / ||G||', which is confusing because G appears on both sides with different meanings. Please use distinct symbols for the raw Gaussian kernel and the normalized vignette mask, and clarify how the Gaussian kernel is computed from the image dimensions.","section":"Appendix A.3, Eq. (A.3)"},{"comment":"For the URPC entry, the dataset column lists five separate counts in parentheses, one per year/version, but the 'Image Number' column is empty; this is inconsistent with other rows and should be clarified either by giving a single aggregate count or by splitting the row by version.","section":"Table 1"},{"comment":"There are minor typographical issues, including 'Moreso' for 'More so,' 'Incase' for 'In case,' and inconsistent capitalization of 'scallop' in Figure 8 and Table 7. A careful proofreading pass would resolve these.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The literature review portion is competent and the authors are honest about limitations, which is a positive sign. However, the two case studies are currently pilot-scale experiments whose quantitative conclusions are not supported by the reported evidence, and at least one conclusion is internally contradicted by the authors' own discussion. I do not recommend rejection because the issues are fixable within the manuscript's scope: additional seeded runs, a matched-size ablation, tempering the language of the conclusions, and reporting or removing the PSNR/SSIM claim would address the load-bearing weaknesses. The journal should also consider whether the case studies belong in a review venue or whether they should be reported and peer-reviewed as a standalone empirical paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nWhat you should know: this is a structured review of underwater object detection with two original case studies bolted on. The review portion is competent and honest, but the empirical claims in the case studies are not supported by the numbers reported.\n\nWhat's actually new: the five-category taxonomy is a re-organization of known material, not a new framing. The genuinely new bits are (1) a DALL-E 3 synthetic augmentation experiment with YOLO11 on RF100, and (2) a LoRA fine-tune of Florence-2 for underwater detection. The review of LVLMs in UOD is timely, and the discussion of hallucination in Florence-2 is honest.\n\nWhat the paper does well: the survey chapters are well-organized, the citations look solid, and the limitations paragraphs are unusually candid. The authors explicitly admit the synthetic data gain is small and possibly insufficient, and they admit the Florence-2 evaluation was impossible due to hallucinated class names. That's good faith.\n\nSoft spots: Table 8 is the load-bearing claim, and it is weak. The differences (mAP@50 0.793 vs 0.796) are within single-run noise; there are no seeds, no error bars, no significance test. Worse, the comparison is confounded: the combined condition changes both data source and dataset size (7,600 to 8,800), so any improvement could just come from having more training data. Their own discussion in Section 5.1.4 acknowledges the gap is small, which undercuts the earlier \"highlight the strength\" phrasing. The Florence-2 case study claims \"strong localization capabilities\" and that the model \"excelled\" at drawing bounding boxes, but the paper itself says hallucinated class names made mAP/recall impossible to compute. That leaves anecdotal evidence. To support the claim they would need a human-evaluation protocol or a metric that tolerates label noise, plus code and prompts.\n\nThe review's self-citations to prior works are fine; those are legitimate reviews. But the PSNR/SSIM claim in Section 5.1.4 has no reported values.\n\nBottom line: as a review, this is usable for newcomers. As an empirical paper, it is a feasibility demo, not a result. The authors are honest about the limits, so it deserves a serious referee—the case studies should be reframed as anecdotal, and Table 8 should be reported as preliminary without causal claims.\n\nRecommendation: send it to peer review with instructions to require major revision on the case-study claims. I would not cite the synthetic-data result, but I might cite the survey if I needed a UOD overview.","headline":"Useful survey with two thin case studies; the quantitative claims about DALL-E 3 augmentation do not survive scrutiny, but the review half is honest and could stand after reframing.","tokens_in":35250,"tokens_out":1771,"would_cite":false,"duration_ms":16491,"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":"A structured review argues that underwater object detection still lacks a complete solution, that DALL-E 3 synthetic images give only marginal YOLO11 gains, and that LVLMs localize well but hallucinate class names.","keywords":["underwater object detection","large vision-language models","synthetic data augmentation","DALL-E 3","Florence-2","LoRA fine-tuning","image enhancement","domain shift"],"falsifier":"Retrain YOLO11 on the original and combined datasets many times with different random seeds and compare the spread of mAP@50 and recall; if the 0.003 and 0.022 gaps overlap across runs, the claimed synthetic-data benefit is not established. Separately, map Florence-2's misspelled class names to ground-truth labels and recompute mAP and recall; if the corrected metrics do not beat a no-fine-tuning baseline, the 'strong localization' claim loses its support.","tokens_in":34129,"feed_emoji":"🌊","tokens_out":8527,"duration_ms":68915,"temperature":0.7,"pith_summary":"The paper is a structured review of underwater object detection (UOD), organized around a five-part taxonomy of challenges: image quality degradation, target-related issues, data-related problems, computational constraints, and detection-methodology limits. Its central argument is that existing UOD methods address parts of these challenges but not the whole, and that large vision-language models (LVLMs) are a promising but unproven direction. To support that argument it runs two case studies: adding 1,200 DALL-E 3 synthetic images to 7,600 real images changes YOLO11 results only slightly (mAP@50 from 0.793 to 0.796, recall from 0.714 to 0.736, precision from 0.805 to 0.780), while a LoRA-fine-tuned Florence-2 draws good bounding boxes yet hallucinates class names so badly that mAP and recall could not be computed. The review's contribution is the taxonomy plus a candid statement of what current LVLMs can and cannot do underwater.","feed_headline":"Synthetic data helps a little; vision-language models hallucinate","feed_subtitle":"1,200 DALL-E 3 images lift YOLO11 recall but not precision; Florence-2 draws boxes yet mangles class names.","key_machinery":"The carrying mechanism is the five-category taxonomy of UOD challenges, which structures the whole review and lets the authors match each challenge family to a solution family (enhancement and restoration, image synthesis, detection architectures, domain adaptation, and efficient fine-tuning). Two case-study pipelines carry the empirical weight: DALL-E 3 text-to-image and image-to-image generation followed by OpenCV enhancement and manual annotation, then YOLO11 training on the original versus combined datasets; and Florence-2 with LoRA applied to attention projection, linear, and convolution layers, fine-tuning 1,929,928 parameters (0.7076% of the model).","core_discovery":"On its own terms, the paper establishes a systematic map of UOD difficulties and shows where the field stands. It claims that the five challenge categories—image quality degradation, target-related issues, data-related challenges, computational and processing constraints, and detection methodology limitations—are best treated together, because solutions aimed at one category (such as enhancement before detection) can interact with another (such as artifacts that hurt small-object localization). The case studies carry the review's main empirical claims: LVLM-generated synthetic data can augment scarce underwater datasets, with Table 8 showing mAP@50 rising from 0.793 to 0.796 and recall from 0.714 to 0.736 at a cost in precision (0.805 to 0.780); and parameter-efficient LoRA fine-tuning of Florence-2 enables strong bounding-box localization for small underwater objects while producing misspelled, ungrounded class names that make standard metrics meaningless. The paper therefore presents LVLMs as a direction with demonstrated localization promise and equally demonstrated output-reliability problems, not as a finished solution.","pith_inferences":["One testable reading of Table 8 is that the synthetic images teach YOLO11 to propose more candidate boxes (higher recall) at the cost of false positives (lower precision); per-class precision-recall curves for the four target classes would show whether the gain is concentrated in classes that are rare in the real data.","Because only single training runs are reported, the 0.003 mAP@50 gain could easily be seed noise; repeating both training conditions with several seeds and holding the real-image count fixed would settle whether the synthetic set itself, rather than simply more training images, drives the change.","Florence-2's misspelled labels look like a lexical or tokenization failure rather than a visual grounding failure, since localization is reportedly strong; a constrained decoding or label-mapping postprocessor could recover usable metrics and clarify whether the bottleneck is semantic hallucination or vocabulary coverage.","The taxonomy could double as an evaluation rubric for future UOD papers, requiring authors to state which challenge categories a method attacks and which it leaves untouched, which would make cross-paper comparison less fragmented."],"forward_implications":["If LVLM synthetic data is used to augment small underwater datasets, the expected effect is modest: Table 8 shows mAP@50 moving from 0.793 to 0.796 and recall from 0.714 to 0.736, with precision dropping from 0.805 to 0.780.","LoRA fine-tuning of Florence-2 on 8,800 images can produce strong bounding-box localization for small objects while updating only 0.7076% of the model parameters, making LVLM adaptation computationally feasible.","Class-name hallucination is the binding constraint for LVLM-based UOD: as long as outputs like 'echinullop' and 'starchin' replace correct labels, standard mAP and recall cannot be computed and practical deployment is blocked.","Real-time UOD with LVLMs remains unresolved; the paper points to efficient fine-tuning, prompt tuning, and lightweight architectures as the next steps rather than a demonstrated result.","The five-category taxonomy gives a shared vocabulary for matching a specific underwater failure (turbidity, small objects, class imbalance, latency, bounding-box overlap) to the solution families reviewed."],"supporting_citations":[{"why":"Supplies the baseline account of underwater imaging challenges and dataset limitations that the review's taxonomy extends.","marker":"[5]"},{"why":"The DUO dataset provides the example images used to illustrate each challenge category.","marker":"[17]"},{"why":"The public underwater dataset whose 7,600 real images anchor the synthetic-augmentation case study.","marker":"[31]"},{"why":"Florence-2 is the LVLM being adapted in the LoRA fine-tuning case study.","marker":"[122]"},{"why":"Supplies the LoRA method that lets fine-tuning update only 0.7% of Florence-2's parameters.","marker":"[138]"},{"why":"Provides the quality-diversity-complexity framework used to interpret whether DALL-E 3 synthetic data helps detection.","marker":"[137]"},{"why":"Establishes the LVLM background that motivates using vision-language models for underwater detection.","marker":"[6]"}],"fun_headline_variants":["Underwater detection review: five challenge areas, one roadmap","Vision-language models: great boxes, garbled labels","Synthetic data: tiny gains, real cost","LVLMs show promise underwater but can't name what they see","From traditional to LVLMs: a critical map of underwater detection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The case-study conclusions assume that the small Table 8 improvements from adding synthetic images are real effects of the data rather than run-to-run noise, and that Florence-2's qualitative bounding boxes demonstrate localization strength despite hallucinated class labels.","fun_headline_variants_meta":{"raw":{"variants":["Underwater detection review: five challenge areas, one roadmap","Vision-language models: great boxes, garbled labels","Synthetic data: tiny gains, real cost","LVLMs show promise underwater but can't name what they see","From traditional to LVLMs: a critical map of underwater detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000415,"raw_usage":{"total_tokens":2176,"prompt_tokens":1013,"completion_tokens":1163,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":629,"completion_tokens_details":{"reasoning_tokens":1081}},"tokens_in":629,"tokens_out":1163,"duration_ms":10244,"temperature":1.0,"reasoning_tokens":1081,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:01:12.894000+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain YOLO11 on the original and combined datasets many times with different random seeds and compare the spread of mAP@50 and recall; if the 0.003 and 0.022 gaps overlap across runs, the claimed synthetic-data benefit is not established. Separately, map Florence-2's misspelled class names to ground-truth labels and recompute mAP and recall; if the corrected metrics do not beat a no-fine-tuning baseline, the 'strong localization' claim loses its support.","supporting_citations":[],"review_version":2}