{"id":"0b745a95-ff8d-4d36-a10b-13c4293998f1","arxiv_id":"2411.11477","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"SL-YOLO reports better small-object detection on VisDrone2019 than YOLOv8s, mostly by adding a P2 detection head and standard lightweight modules.","lead":"SL-YOLO, a modified YOLOv8 model, adds a small-target detection layer, a tweaked feature-fusion network, and lighter convolution modules for drone imagery. On VisDrone2019 it reports higher mAP (46.9% vs 43.0%) with fewer parameters (9.6M vs 11.1M) than the YOLOv8s baseline.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Ablation evidence for the proposed HEPAN/C2fDCB/SCDown components rests on single-run deltas of 0.1-1.0 mAP on VisDrone-val; with no error bars, the final accuracy gain may reduce to the known P2 head.","rationale":"The reader's weakest assumption — that single-run results make 0.1-0.3 point differences meaningless — is correct and is the core of this attack. I extend it by showing exactly where it bites: Table 3 attributes the whole 3.9-point improvement to the P2 head plus small deltas from the proposed modules, so without variance estimates the novelty of HEPAN, C2fDCB, and SCDown is unverifiable. I also flag the internal tension that the lightweight modules reduce accuracy and that GFLOPs rise to 36.7 versus 28.5 for YOLOv8s, undercutting the 'lighter' rhetoric. These are correctness and calibration concerns, not accusations of fraud. The quantified headline (46.9 mAP, 9.6M params) is not contradicted by the paper's own data, so a conditional acceptance asking for seeds, code, and tempered claims remains appropriate. I do not see grounds to reject outright.","tokens_in":11090,"tokens_out":10846,"duration_ms":102918,"concrete_test":"Train YOLOv8s-P2, YOLOv8s-P2+HEPAN, and final SL-YOLO on VisDrone2019 with at least three random seeds each under the Section 4.2 protocol, and report mean plus or minus standard deviation of mAP@0.5 and mAP@0.5:0.95. Also report torchinfo parameter counts for each row. If the HEPAN-vs-P2 and final-vs-P2 mAP differences fall within one standard deviation, or if the parameter count increases when the P2 head is added, the proposed components' accuracy contributions are unsupported and the headline reduces to the known P2-head modification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim is that SL-YOLO's proposed components, not just a P2 head, make it stronger and lighter. Table 3 shows the entire 3.9-point mAP@0.5 gain over YOLOv8s appears after adding the P2 small-object head (43.0 to 46.2); HEPAN adds +1.0 (46.2 to 47.2), and the two lightweight modules then reduce accuracy by 0.2 and 0.1 points to a final 46.9. Section 4.2 reports no seeds, error bars, or statistical tests, so the +1.0 HEPAN gain and the -0.3 total loss from C2fDCB plus SCDown are indistinguishable from run-to-run variation on a 548-image validation set. The abstract's 'without sacrificing detection capabilities' is also contradicted by these point estimates. Consequently, the only robust part of the headline is the deterministic parameter decrease (11.1M to 9.6M); the claimed contributions of the three proposed modules to accuracy are not established by the evidence presented. The suspicious parameter drop from 11.1M to 10.6M when 'adding' the P2 head further suggests the ablation rows differ in more than the variable being tested.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SL-YOLO, an enhancement of YOLOv8s for small-object detection in drone imagery. The proposed architecture adds a P2 small-target detection head, a Hierarchical Extended Path Aggregation Network (HEPAN), and two lightweight modules (C2fDCB and SCDown). On VisDrone2019-val, the authors report mAP@0.5 improving from 43.0% to 46.9% and mAP@0.5:0.95 from 26.0% to 28.9%, while parameters drop from 11.1M to 9.6M, at 132 FPS. The paper also reports comparisons on VisDrone2019-test and an ablation study.","tokens_in":11248,"tokens_out":2933,"duration_ms":28059,"significance":"If the reported gains were robust, SL-YOLO would offer a modest accuracy/size trade-off for drone detection, which is a practically relevant setting. The paper provides a useful empirical comparison across several YOLO variants on both val and test splits, and it reports parameters, GFLOPs, and FPS. However, the central evidence is a set of single-run point estimates with no error bars, and the ablation table contains an internal inconsistency in parameter counts. In its current form, the manuscript does not establish that the proposed components make a reliable contribution beyond the known P2-head modification.","major_comments":[{"comment":"All reported metrics come from a single training run per configuration, with no seeds, error bars, or statistical tests. The ablation deltas that support the paper's central claims are within typical run-to-run variation on a 548-image validation set: HEPAN gives +1.0 mAP@0.5 (46.2 to 47.2), C2fDCB gives -0.2, and SCDown gives -0.1. Without variance estimates, the statement in the abstract that SL-YOLO achieves a 'significant improvement' is not justified. Please report mean and standard deviation over at least 3-5 seeds, or provide a statistical test, for the main comparisons and the full ablation.","section":"Section 4.2 and Table 3"},{"comment":"Adding a P2 small-target head to YOLOv8s is reported to reduce parameters from 11.1M to 10.6M while increasing GFLOPs from 28.5 to 36.7. Since a P2 head adds convolutional layers, this decrease is inconsistent with the stated modification and suggests that the ablation rows differ in more than the variable being tested. Please provide the exact module configuration and parameter accounting for each row, and explain why the parameter count decreases when the P2 head is added.","section":"Table 3, row 2"},{"comment":"The paper claims the lightweight modules C2fDCB and SCDown reduce model size 'without sacrificing detection capabilities.' The point estimates in Table 3 contradict this: C2fDCB lowers mAP@0.5 by 0.2 and SCDown by another 0.1, and Tables 5 and 6 state that C2f and Conv are respectively slightly better than C2fDCB and SCDown. The claim should be rephrased as an accuracy loss that may be within uncertainty, or the authors should provide evidence that the differences are not statistically meaningful.","section":"Abstract, Introduction, and Tables 3, 5, 6"}],"minor_comments":[{"comment":"The abstract and introduction use promotional language such as 'revolutionary,' 'pioneering,' 'unparalleled,' and 'break the bottleneck'; these terms are not supported by the modest measured improvements and should be replaced with precise, quantitative descriptions.","section":"Abstract and Introduction"},{"comment":"The text attributes BiFPN to 'Li [30]', but reference [30] is the EfficientDet paper by Tan et al.; please correct the citation.","section":"Section 3.2"},{"comment":"The FPS measurement is described as using 'a single image' and bs=1, but no details are given about warm-up, number of repetitions, or whether the reported FPS is the mean of multiple runs; please describe the protocol and report variance.","section":"Section 4.2"},{"comment":"The term 'RepVGGDW convolution' is used but never defined or referenced; please clarify whether this is a standard RepVGG block or a new module, and show its structure.","section":"Section 3.3.1"},{"comment":"There are numerous typos and grammatical issues, for example 'we designs the C2fDCB' in the Introduction, 'UA V aerial images' in Section 3.1, and 'also be introduced' in Section 3.2; a thorough proofread is needed.","section":"Throughout"},{"comment":"The sentence comparing SL-YOLO and YOLOv8s-p2 says 'higher accuracy (37.4% vs 38.3%)' but the order of the numbers is confusing; it should read '38.3% vs 37.4%'.","section":"Section 4.3, Table 2"}],"recommendation":"major_revision","confidential_remarks":"The paper is a straightforward architecture-engineering contribution with single-run empirical evidence. The central problem is not that the architecture is hopeless, but that the reported ablation does not currently isolate the proposed contributions, and the parameter-count inconsistency in Table 3 undermines confidence in the configuration control. I would ask the authors for multi-seed experiments, corrected parameter accounting, and a revised abstract that matches the evidence. If these are provided, the paper could be publishable as a modest systems paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a modest, honestly-reported architecture tweak. The paper adds a P2 small-object head, a PANet/BiFPN variant (HEPAN), and two lightweight modules to YOLOv8s, and reports 46.9 vs 43.0 mAP@0.5 on VisDrone2019-val with fewer parameters. The ablation shows the P2 head contributes +3.2 points, HEPAN another +1.0, and then C2fDCB (-0.2) and SCDown (-0.1) each shave off a little accuracy. So the headline \"stronger and lighter without sacrificing detection capabilities\" is not supported by the paper's own numbers—the lightweighting costs a small but consistent amount of mAP.\n\nWhat the paper does well: the comparison tables cover a broad set of YOLO versions on both val and test splits, and the ablation is transparent about the degradation. The parameter counts are deterministic and the setup is standard, which makes the main result reproducible in principle. The writing is clear, and the related work is appropriately positioned.\n\nWhere it's soft: the single-run results have no seeds, error bars, or statistical tests. On a 548-image validation set, the HEPAN gain of +1.0 mAP and the 0.2/0.1 losses from the lightweight modules are indistinguishable from run-to-run noise. Also, the parameter count drops from 11.1M to 10.6M when \"adding\" the P2 head, which is suspicious—adding a head should add parameters, so the baseline and the P2 variant likely differ in more than just the head. The abstract's language (\"revolutionary,\" \"pioneering,\" \"unparalleled\") is overblown, and SCDown is taken from YOLOv10 without explicit attribution in the methodology section, though the citation is present elsewhere.\n\nThe central empirical claim—that the full model beats YOLOv8s by a few points at lower parameter count—is probably true, but the component-level contributions are not established. This is a useful data point for practitioners working on drone detection, not a research advance.\n\nRecommendation: it deserves a serious referee if the venue tolerates incremental architecture papers, but only with major revisions: release code, run multiple seeds with variance reported, fix the baseline confusion, and temper the claims. As-is, it's a conditional accept at best, and a desk reject would also be defensible on novelty and rigor grounds.","headline":"Incremental YOLOv8 variant whose P2 head explains most of the gain; the lightweight modules slightly hurt accuracy, and single-run results don't support component-level claims.","tokens_in":11893,"tokens_out":2084,"would_cite":false,"duration_ms":22634,"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":"SL-YOLO lifts drone small-target mAP from 43.0% to 46.9% on VisDrone2019 while cutting parameters from 11.1M to 9.6M.","keywords":["drone target detection","small object detection","YOLOv8","HEPAN","feature pyramid network","lightweight network","VisDrone2019","depthwise separable convolution"],"falsifier":"Train the models in Table 3 with at least five random seeds and report mean plus standard deviation for each. If the mAP@0.5 difference between YOLOv8s and SL-YOLO, or between adjacent ablation rows, falls within one standard deviation, the claimed 'stronger and lighter' advantage and the step-by-step ablation do not survive.","tokens_in":10793,"feed_emoji":"🚁","tokens_out":5074,"duration_ms":41566,"temperature":0.7,"pith_summary":"This paper argues that drone-view small-target detection can be materially improved without needing a larger model. On the VisDrone2019 benchmark, the proposed SL-YOLO model raises mean average precision at 0.5 IoU from 43.0% (YOLOv8s) to 46.9% while cutting parameters from 11.1M to 9.6M. The gains come from an added high-resolution detection head, a modified feature-fusion neck called HEPAN, and two lightweight modules. The paper positions this result as useful for real-time drone monitoring in resource-constrained settings.","feed_headline":"Lighter drone detector lifts small-target mAP from 43.0% to 46.9%","feed_subtitle":"Adds a high-resolution head and lighter neck to YOLOv8s while cutting parameters from 11.1M to 9.6M.","key_machinery":"The central machinery is the P2 small-target detection head plus the Hierarchical Extended Path Aggregation Network (HEPAN), a neck that adds extra convolutional layers and residual connections between the top-down and bottom-up fusion paths, generating a 160×160 feature map that preserves small-object information usually lost in downsampling. Alongside it, the C2fDCB module (a C2f block rebuilt with depthwise separable convolutions and RepVGGDW reparameterization) and the SCDown module (1×1 channel compression followed by strided depthwise convolution) reduce parameter count and FLOPs. The argument runs through the ablation table, where each component is added to the YOLOv8s baseline and its contribution to accuracy and cost is read off.","core_discovery":"SL-YOLO is a YOLOv8s variant that adds a P2-level detection head producing a 160×160 feature map, replaces the PAN neck with HEPAN (which inserts extra convolutional layers and residual connections into the fusion path), and swaps the C2f block for C2fDCB (depthwise separable convolution combined with RepVGG-style reparameterization), with SCDown handling downsampling. On VisDrone2019-val the full model reaches mAP@0.5 of 46.9% and mAP@0.5:0.95 of 28.9%, up from 43.0% and 26.0% for baseline YOLOv8s, while parameters drop from 11.1M to 9.6M and inference runs at 132 FPS. The ablation attributes most of the gain to the P2 head (+3.2 mAP), with HEPAN adding another +1.0; the lightweight modules trade about 0.3 mAP for a parameter reduction of roughly 1.7M.","pith_inferences":["If the P2 head is the dominant contributor, then a multi-seed comparison of SL-YOLO against YOLOv8s-p2 (which scores 46.2 mAP at the same 36.7 GFLOPs) would isolate whether HEPAN, C2fDCB, and SCDown add a real edge beyond the head alone.","The paper's own data suggest a cost-accuracy frontier: the HEPAN-only configuration reaches 47.2 mAP at 11.3M parameters, while the full SL-YOLO trades 0.3 mAP for a 1.7M parameter cut, meaning a practitioner could choose either point depending on memory constraints.","Because the ablation appears to rely on a single run per configuration, the 0.1-point drops from C2fDCB and SCDown fall in the range where training noise matters; reporting variance over several seeds would clarify whether the lightweight modules are truly lossless."],"forward_implications":["A P2-style detection head alone accounts for most of the small-target gain, so other detector architectures could adopt it without the rest of the SL-YOLO stack.","HEPAN's residual-enhanced fusion adds roughly one mAP point on top of the P2 head, suggesting deeper neck fusion helps small objects at a modest cost (+0.7M parameters relative to P2-only).","The C2fDCB and SCDown modules together cut 1.7M parameters and 1.4 GFLOPs while losing only 0.3 mAP, a favorable trade for edge deployment.","At 9.6M parameters and 132 FPS, SL-YOLO offers a real-time profile comparable to YOLOv8s while outperforming it by 3.9 mAP@0.5 on VisDrone2019-val.","On VisDrone2019-test the model trails larger YOLOv9m and YOLOv11m by about 0.5 mAP, so the accuracy claim is strongest against similar-size YOLOv8s models."],"supporting_citations":[{"why":"Provides the VisDrone2019 dataset used for all training and evaluation, including the validation and test splits.","marker":"[4]"},{"why":"Supplies the YOLOv8s baseline model and the official code base that SL-YOLO modifies.","marker":"[16]"},{"why":"Introduces depthwise separable convolutions, the basis of the C2fDCB block's parameter reduction.","marker":"[2]"},{"why":"Introduces RepVGG reparameterization, used in the RepVGGDW convolution within C2fDCB.","marker":"[3]"},{"why":"Describes the PANet neck structure that serves as the baseline and that HEPAN extends.","marker":"[20]"},{"why":"Defines BiFPN, the feature-fusion structure compared against HEPAN in the network-structure analysis.","marker":"[30]"}],"fun_headline_variants":["Drone detector mAP jumps 3.9 points with lighter YOLO","SL-YOLO: stronger and lighter, +3.9 mAP on VisDrone","Lighter YOLO for drones: 46.9 mAP, 9.6M params, 132 FPS","Small-target boost: stronger YOLO, fewer parameters","YOLOv8s variant: +3.9 mAP, -1.5M params, 132 FPS"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Each configuration in the ablation was trained once, so reported gaps of 0.1 to 0.3 mAP points between similar models may be run-to-run noise rather than real architectural differences.","fun_headline_variants_meta":{"raw":{"variants":["Drone detector mAP jumps 3.9 points with lighter YOLO","SL-YOLO: stronger and lighter, +3.9 mAP on VisDrone","Lighter YOLO for drones: 46.9 mAP, 9.6M params, 132 FPS","Small-target boost: stronger YOLO, fewer parameters","YOLOv8s variant: +3.9 mAP, -1.5M params, 132 FPS"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000769,"raw_usage":{"total_tokens":3452,"prompt_tokens":1037,"completion_tokens":2415,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":653,"completion_tokens_details":{"reasoning_tokens":2292}},"tokens_in":653,"tokens_out":2415,"duration_ms":18831,"temperature":1.0,"reasoning_tokens":2292,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T18:27:41.598065+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the models in Table 3 with at least five random seeds and report mean plus standard deviation for each. If the mAP@0.5 difference between YOLOv8s and SL-YOLO, or between adjacent ablation rows, falls within one standard deviation, the claimed 'stronger and lighter' advantage and the step-by-step ablation do not survive.","supporting_citations":[{"cited_title":"Visdrone-det2019: The vision meets drone ob- ject detection in image challenge results","cited_arxiv_id":null,"evidence_quote":"Provides the VisDrone2019 dataset used for all training and evaluation, including the validation and test splits."},{"cited_title":"Ultralytics yolo","cited_arxiv_id":null,"evidence_quote":"Supplies the YOLOv8s baseline model and the official code base that SL-YOLO modifies."},{"cited_title":"Xception: Deep learning with depthwise separable convolutions","cited_arxiv_id":null,"evidence_quote":"Introduces depthwise separable convolutions, the basis of the C2fDCB block's parameter reduction."},{"cited_title":"Repvgg: Making vgg-style convnets great again","cited_arxiv_id":null,"evidence_quote":"Introduces RepVGG reparameterization, used in the RepVGGDW convolution within C2fDCB."},{"cited_title":"Path aggregation network for instance segmentation","cited_arxiv_id":null,"evidence_quote":"Describes the PANet neck structure that serves as the baseline and that HEPAN extends."},{"cited_title":"Efficient- det: Scalable and efficient object detection","cited_arxiv_id":null,"evidence_quote":"Defines BiFPN, the feature-fusion structure compared against HEPAN in the network-structure analysis."}],"review_version":1}