{"id":"7422ab61-85cb-4812-b9be-d0348cc18cf7","arxiv_id":"2501.12169","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"SVGS-DSGAT, a combination of GraphSage, saliency-guided visual attention, and difference-similarity graph attention, is claimed to outperform existing detectors on two underwater datasets, but the supporting evidence is inconsistent and unreproducible.","lead":"This paper proposes an underwater object detection model called SVGS-DSGAT that combines GraphSage, a saliency attention module, and a difference-similarity graph attention module, reporting mAP of 40.8% on URPC 2020 and 41.5% on SeaDronesSee. A general reader might care because robust underwater perception matters for ocean monitoring, but the paper's evidence is weakened by missing code, missing data, and internal inconsistencies.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed mAP gains are not attributable to the proposed architecture because the image-to-graph construction is never specified, and the reported ablation and parameter tables are internally inconsistent.","rationale":"The reader's weakest-assumption analysis identified exactly the load-bearing gap: the graph construction is never defined, so the described architecture cannot be instantiated from the paper. My stress-test pass confirms this and adds that even the empirical reporting is internally inconsistent, which makes the claimed performance unattributable to the proposed modules. I agree with the REJECT verdict: the central claim is not checkable as presented. This is not an objection to the general idea of combining graph sampling and attention for underwater detection; it is a correctness and reproducibility objection. The paper provides no code, no data, no formal verification, and no consistent ablation reporting, so accepting the headline mAP numbers would require trusting unspecified implementation choices. The concrete test I propose would settle the concern: if the authors can supply a precise graph-construction definition and code that reproduces Table 6, the claim could be re-evaluated; until then, the rejection stands.","tokens_in":17819,"tokens_out":3631,"duration_ms":39289,"concrete_test":"Obtain from the authors a complete graph-construction specification and executable code: for a 512x512 URPC 2020 image, state precisely what each node is (superpixel, grid patch, object proposal, or other), how edges are formed (k-NN, spatial adjacency, or other), and what the feature matrix X contains. Then run the released code with the stated hyperparameters (Table 2) on URPC 2020 test-A and recompute Table 6's 'Ours' row. If the published mAP of 40.8 is not reproduced, or if the graph construction cannot be specified without inventing details absent from the paper, the claimed superiority of SVGS-DSGAT is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that SVGS-DSGAT outperforms mainstream detectors on URPC 2020 and SeaDronesSee. For that claim to be checkable, the architecture must be fully specified. It is not. Section 3.2 states that preprocessed underwater image data is input into GraphSage, and Algorithm 1 takes a graph G=(V,E) and feature matrix X as input, but the paper never defines how an image is converted into nodes V, edges E, or node features X. Section 3.4's DSGAT similarly assumes nodes and neighborhoods N(i) without specifying their origin. Consequently, Equations (1)-(9) cannot be evaluated on an actual underwater image from the information given, and the reported 40.8% / 41.5% mAP cannot be traced to the described GraphSage+SVAM+DSGAT pipeline. This is not a stylistic gap: without a node/edge definition, the model is not reproducible from the text, and the performance could come from an unspecified backbone, postprocessing, or a different graph construction. The concern is reinforced by internal numerical inconsistencies: Section 4.7 says the complete model achieves 41.5% mAP on URPC 2020 and 42.3% on SeaDronesSee, while Tables 3, 4, and 6 list Ours as 40.8% and 41.5%; the same text says GraphSage+SVAM improves mAP by 3.7%, while Table 6 shows a 9.6-point gain over GraphSage alone; Table 2 reports 3,768,945 parameters, but Table 5 reports 6.33M for Ours; Table 2 sets 200 training epochs, but Section 4.8 says 300 epochs. The claimed statistical significance (p<0.05) is stated without any test procedure, sample size, or variance estimate. Together these issues mean the central empirical claim is currently unverifiable and internally unchecked.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SVGS-DSGAT, a model that combines GraphSage, a Saliency-Guided Visual Attention Module (SVAM), and a Difference Similarity Graph Attention Module (DSGAT) for underwater object detection, with an IoT-enabled system architecture. The authors report mAP of 40.8% on URPC 2020 and 41.5% on SeaDronesSee, claim statistical significance with p<0.05, and present comparisons against YOLO-family and other detectors, plus ablations and parameter-efficiency numbers. The central claim is that this combination outperforms mainstream detectors at lower parameter count and computational cost.","tokens_in":18265,"tokens_out":3226,"duration_ms":28976,"significance":"If the claimed results were reliably established, the paper would offer a useful data point on combining graph-based feature aggregation with saliency-guided and difference-similarity attention for underwater detection, and the IoT framing could be relevant for deployment-oriented research. The paper includes experiments on two public datasets, comparisons with several baselines, and an ablation study, which are appropriate ingredients for an empirical contribution. However, the significance is severely undercut by the lack of a precise definition of the graph construction from images, by multiple internal numerical contradictions, and by an unsupported statistical-significance claim. As it stands, the reported performance cannot be attributed to the described architecture, and the paper does not provide a reproducible or verifiable method.","major_comments":[{"comment":"The paper never defines how an underwater image is converted into the graph G=(V,E) and feature matrix X that Algorithm 1 and the DSGAT equations require. Section 3.2 only states that \"preprocessed underwater image data is input into the GraphSage module,\" and Section 3.4 assumes nodes, neighborhoods N(i), and features h_i without explaining their origin. Without a node/edge/feature definition, Equations (1)–(9) cannot be evaluated on an actual image, and the architecture is not reproducible from the text. This is a load-bearing gap: the claimed mAP gains cannot be traced to the described GraphSage+SVAM+DSGAT pipeline.","section":"§3.2, Algorithm 1, §3.4 (Eqs. 6–9)"},{"comment":"The reported performance numbers are internally inconsistent. The abstract, Section 4.5, and Tables 3/4/6 list the full model as 40.8% mAP on URPC 2020 and 41.5% on SeaDronesSee, but Section 4.7 states that \"the complete model proposed in this study (Ours) achieves ... an mAP of 41.5% on the URPC 2020 dataset and 42.3% on the SeaDronesSee dataset.\" These are different values for the same configuration, and the manuscript does not explain which is correct or how they were computed.","section":"§4.5 vs §4.7 vs Tables 3, 4, 6"},{"comment":"Additional numerical contradictions affect the model's reported efficiency and training setup. Table 2 lists \"Model Parameters\" as 3,768,945 (about 3.77M) and \"Training Epochs\" as 200, while Table 5 reports the proposed model's parameter count as 6.33M, and Section 4.8 states \"The model was trained for 300 epochs.\" These discrepancies prevent the reader from knowing the actual parameter count and training configuration, undermining the efficiency claims and the reproducibility of the experiments.","section":"Table 2 vs Table 5, §4.8"},{"comment":"The ablation discussion is internally inconsistent with the table it refers to. The text says \"The combination of GraphSage and SVAM ... with mAP increasing by 3.7% and 3.8%,\" but Table 6 shows GraphSage+SVAM at 37.9% versus GraphSage alone at 28.3%, an increase of 9.6 points; the text itself later acknowledges this 9.6-point improvement. The sentence \"after adding SVAM, these figures increase to 36.2% and 37.0% respectively; further adding DSGAT raises these to 32.9% and 33.7%\" is also contradictory, since 32.9/33.7 correspond to DSGAT alone in Table 6, not to GraphSage+SVAM+DSGAT, which is listed as 40.8/41.5.","section":"§4.7, Table 6"},{"comment":"The paper states that \"p-values below 0.05\" indicate statistical significance across metrics, but no statistical test, number of repeated runs, variance measures, or error bars are reported. Without this information, the claim is unsupported and cannot be independently checked, especially given the numerical inconsistencies noted above.","section":"§4.5, statistical significance claim"}],"minor_comments":[{"comment":"The paragraph claiming that \"GraphSage contributes to a 12% increase in accuracy, SVAM improves precision by 8%, and DSGAT provides a 10% improvement in overall robustness\" presents numbers that do not appear in Table 6 and are not derived from any described procedure; these figures should be removed or substantiated.","section":"§4.7, final paragraph"},{"comment":"The row \"Number of Layers 200\" is unclear: this seems to conflate training epochs with network depth, and the abbreviation is not consistent with the rest of the paper.","section":"Table 2"},{"comment":"The text refers to \"Mask IoU thresholds\" in the discussion of AP/ΔAP curves, but the object detection experiments do not involve masks; this terminology appears to be a leftover from a segmentation context and should be corrected.","section":"Figure 6 and §4.5"},{"comment":"Several references are missing complete metadata (e.g., some entries lack page numbers or publisher information), and the citation keys in the text sometimes carry suffixes (e.g., \"Lo, Layeghy, Sarhan, Gallagher and Portmann (2022)\") that are not reflected consistently in the reference list.","section":"References"}],"recommendation":"reject","confidential_remarks":"The manuscript combines existing modules (GraphSage, SVAM from Islam et al., DSGAT adapted from Lian et al.) but does not provide the crucial graph-construction specification that would make the combination a testable new method. The internal numerical contradictions (abstract vs. §4.7, Table 2 vs. Table 5, epochs 200 vs. 300) and the unsupported statistical claim are not local presentation issues; they affect the central claims of performance and efficiency. The IoT component is described only narratively and is not experimentally validated, so it adds little beyond framing. These issues would require a substantial rewrite and new experiments to address, which is beyond the scope of a minor or moderate revision for this venue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Let me save you a close read: the central claim is unverifiable as written. The paper reports mAP numbers for a GraphSage+SVAM+DSGAT detector on URPC 2020 and SeaDronesSee, but it never defines how an image becomes a graph—no node construction, no edge rule, no feature matrix. Algorithm 1 starts from G=(V,E), and Section 3.2 just says preprocessed images are input into GraphSage. That gap alone makes the architecture unreproducible from the text.\n\nThere is something worth acknowledging: combining those three published modules is a plausible direction, and the paper does run a full set of comparisons and ablations on two real benchmarks. The cited sources for SVAM and DSGAT are legitimate. If the composition worked, a 0.1-point mAP edge over YOLOv8 on URPC and 1.9 on SeaDronesSee with fewer parameters would be a modest but legitimate incremental result.\n\nBut the soft spots are load-bearing. The reported numbers contradict each other: the abstract says 40.8/41.5, Section 4.7 says 41.5/42.3. Table 2 lists 3.77M parameters and 200 epochs; Table 5 says 6.33M; Section 4.8 says 300 epochs. The ablation text says GraphSage+SVAM gains 3.7%, while Table 6 shows a 9.6-point jump over GraphSage alone. The p-value claim has no test procedure or variance estimate. These are not cosmetic typos; they mean the reader cannot tell which numbers are real.\n\nThere is also a category error: Figure 6 plots AP against 'Mask IoU thresholds' for a detection model, which indicates the evaluation description does not match the task. The IoT integration, despite being in the title, is only a system diagram; there is no IoT experiment.\n\nWho gets value from this? Possibly someone mining references for combinations of attention and GNN modules. But as a research paper, it is not ready for peer review. The empirical claim cannot be trusted, and the architecture cannot be reproduced. I'd recommend desk rejection. If the authors supply code, define the graph construction, and reconcile the numbers, there might be a salvageable incremental result, but the current version is not a serious submission.","headline":"The paper's empirical claim is untraceable to its architecture: the image-to-graph mapping is undefined and the reported mAP, parameter, and epoch numbers conflict internally, so desk rejection is the right call.","tokens_in":18799,"tokens_out":4533,"would_cite":false,"duration_ms":40419,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"SVGS-DSGAT reports 40.8% and 41.5% mAP on underwater benchmarks, beating YOLO-family baselines with fewer parameters.","keywords":["underwater object detection","Internet of Things","GraphSage","SVAM","DSGAT","graph neural network","attention mechanism","YOLO comparison"],"falsifier":"The decisive check is to re-run the model on URPC 2020 and SeaDronesSee with the image-to-graph mapping specified in enough detail to reproduce it; if changing the node or edge definition shifts mAP by several points, the reported gains belong to the graph construction rather than to the SVGS-DSGAT modules. A second check available from the paper itself is the ablation narrative: the text says adding DSGAT to GraphSage+SVAM raises mAP to 32.9 and 33.7, while Table 6 lists 40.8 and 41.5 for the full model, so the numbers as printed leave the incremental contribution of DSGAT ambiguous.","tokens_in":17621,"feed_emoji":"🌊","tokens_out":9169,"duration_ms":79876,"temperature":0.7,"pith_summary":"The paper proposes SVGS-DSGAT, an underwater object detector that combines GraphSage (graph sampling and aggregation), SVAM (saliency-guided visual attention), and DSGAT (difference-similarity graph attention), together with an IoT layer for data transmission. Its central claim is that this stack reaches 40.8% mean average precision on URPC 2020 and 41.5% on SeaDronesSee, outperforming YOLOv4 through YOLOv8, Faster R-CNN, SSD, EfficientDet, RetinaNet, and CenterNet while using about 6.33 million parameters and 10.28 billion FLOPs. The authors attribute the gain to graph-based structure capture plus two complementary attention mechanisms that cope with high noise, low contrast, and small targets. A sympathetic reader would care because underwater robot vision is limited by exactly those conditions, and the reported accuracy comes at a lower computational cost, which matters for real-time deployment.","feed_headline":"Underwater detector beats YOLO models with fewer parameters","feed_subtitle":"SVGS-DSGAT reports 40.8% mAP on URPC 2020 and 41.5% on SeaDronesSee using graph sampling plus dual attention.","key_machinery":"The central object is the SVGS-DSGAT module stack. GraphSage samples a fixed-size set of neighbors for each node and aggregates their features to build node embeddings, which is how the image is treated as graph-structured data. SVAM computes a saliency map from pixel intensities, uses it to build an attention-weighted feature map, applies edge enhancement through gradients and a Laplacian, and produces a sigmoid output. DSGAT computes pairwise difference (Euclidean distance in feature space) and similarity (cosine) between node features, combines them into softmax attention weights, and aggregates neighbor features through those weights. The stack converts noisy, low-contrast underwater images into discriminative node embeddings for detection and tracking.","core_discovery":"The authors report that replacing a purely convolutional detector with a graph-sampling backbone followed by saliency-guided and difference-similarity attention improves underwater detection on both accuracy and tracking metrics. On URPC 2020 the model records mAP 40.8%, AP50 75.0%, AP75 78.5%, and MOTA 62.5%; on SeaDronesSee it records mAP 41.5%, AP50 76.2%, AP75 79.0%, and MOTA 63.7%. They attribute the result to GraphSage filtering noise through neighbor sampling and aggregation, SVAM focusing computation on salient regions, and DSGAT setting attention weights from the difference and similarity between node features. The paper also reports p-values below 0.05 for the improvements and an ablation sequence in which the full model outperforms every subset, concluding that each module contributes to robustness in complex underwater environments.","pith_inferences":["The paper never specifies how an underwater image is converted into a graph, so a natural next experiment is to vary the node and edge definitions while holding the rest of the model fixed; the sensitivity of mAP to that choice would reveal how much of the gain is genuinely graph-based.","The IoT component is described at the system level but not isolated experimentally; a testable extension is to measure end-to-end latency and throughput over the LoRa/acoustic link to see whether the benchmark accuracy survives real-time constraints.","If the graph construction is made concrete, the same module stack could be transferred to other low-contrast detection domains, such as aerial search-and-rescue or endoscopic imaging, where saliency and relational attention are also relevant."],"forward_implications":["If the reported numbers hold, graph-plus-attention detection becomes a credible alternative to YOLO-family detectors on underwater benchmarks, improving mAP, AP50, AP75, and MOTA simultaneously.","The lower parameter count (6.33M versus 8.4–8.8M) and FLOPs (10.28B versus 11.5–11.9B) imply the model is cheaper to deploy on resource-limited underwater robots, assuming the IoT transmission layer adds no prohibitive overhead.","The tracking metrics indicate the same features also stabilize multi-object tracking, not just static detection, which matters for robot navigation and monitoring.","The ablation results imply the largest accuracy jump comes from combining GraphSage with SVAM, with DSGAT adding further refinement, so future designs can weigh which module to add first."],"supporting_citations":[{"why":"Supplies the GraphSage graph sampling and aggregation algorithm that the model uses for feature extraction.","marker":"Hajibabaee, Malekzadeh, Heidari, Zad, Uzuner and Jones (2021)"},{"why":"Defines SVAM, the saliency-guided visual attention module that focuses the model on salient image regions.","marker":"Islam, Wang and Sattar (2020)"},{"why":"Provides the difference-similarity graph attention architecture that DSGAT is built on.","marker":"Lian, Li, Cong, Li, Zhang and Kwong (2023)"},{"why":"Provides the URPC 2020 benchmark dataset used for the first evaluation.","marker":"Fu, Liu, Fan, Chen, Fu, Yuan, Zhu and Luo (2023)"},{"why":"Provides the SeaDronesSee benchmark dataset used for the second evaluation.","marker":"Liu, Ma, Xu and Li (2024)"},{"why":"Supplies the YOLOv8 baseline that the model must beat on mAP.","marker":"Qu, Cui, Duan, Lu and Pang (2024)"},{"why":"Supplies the YOLOv4 baseline used in the comparison tables.","marker":"Chen and Fan (2020)"}],"fun_headline_variants":["GraphSage plus dual attention dives past YOLO underwater","IoT graph net beats YOLO in murky-water detection","Underwater robot sees better with graph sampling and attention","Graph backbone lifts underwater detection above YOLO models"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that an underwater image can be turned into a graph with well-defined nodes, edges, and feature vectors, but the paper never specifies that construction, so the reported accuracy cannot be uniquely attributed to the GraphSage/SVAM/DSGAT stack until the mapping is stated.","fun_headline_variants_meta":{"raw":{"variants":["GraphSage plus dual attention dives past YOLO underwater","IoT graph net beats YOLO in murky-water detection","Underwater robot sees better with graph sampling and attention","Graph backbone lifts underwater detection above YOLO models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000181,"raw_usage":{"total_tokens":1302,"prompt_tokens":934,"completion_tokens":368,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":550,"completion_tokens_details":{"reasoning_tokens":300}},"tokens_in":550,"tokens_out":368,"duration_ms":5019,"temperature":1.0,"reasoning_tokens":300,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T17:26:17.719972+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"The decisive check is to re-run the model on URPC 2020 and SeaDronesSee with the image-to-graph mapping specified in enough detail to reproduce it; if changing the node or edge definition shifts mAP by several points, the reported gains belong to the graph construction rather than to the SVGS-DSGAT modules. A second check available from the paper itself is the ablation narrative: the text says adding DSGAT to GraphSage+SVAM raises mAP to 32.9 and 33.7, while Table 6 lists 40.8 and 41.5 for the full model, so the numbers as printed leave the incremental contribution of DSGAT ambiguous.","supporting_citations":[],"review_version":1}