{"id":"cfaa2c6d-55ea-4d0c-8607-4002c36d6457","arxiv_id":"2608.07571","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A survey and benchmark comparison of YOLO-family detectors for UAV traffic monitoring, reporting accuracy and energy-efficiency numbers on VisDrone and a Cyprus aerial vehicle dataset.","lead":"This paper surveys deep-learning vehicle detection from drone footage for traffic monitoring, and adds its own benchmark tests of YOLO variants on two aerial datasets. A generalist might read it for a snapshot of which detection models are accurate versus efficient on drone hardware, though the experimental results need verification.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'highest mAP' claims for YOLOv11m are contradicted by the paper's own Tables III–V, which list an undefined YOLOv26(m) with higher scores; the benchmark conclusion cannot be evaluated until that model is sourced or removed.","rationale":"The reader correctly identified YOLOv26(m) as a load-bearing weakness. My stress-test found an even more direct form of the same problem: the paper's own narrative contradicts its own tables. The text twice asserts YOLOv11m is the best, but Tables III, IV, and V each show YOLOv26(m) with higher mAP values. This is not an external disagreement with consensus; it is an internal inconsistency. If YOLOv26(m) is excluded as unverifiable, the paper should at minimum revise its claims and recompute its rankings; if it is included, the claims as written are false. Either way, the empirical centerpiece does not currently support the stated conclusion. The survey portion appears broadly consistent with known literature and the energy-efficiency derivation (E_f = P_limit/FPS) is arithmetically coherent, but those strengths do not rescue the benchmark claims. A rejection or major revision is warranted, consistent with the reader's verdict. I therefore mark the verdict as unchanged rather than proposing a different outcome.","tokens_in":30951,"tokens_out":2909,"duration_ms":28027,"concrete_test":"Request or locate the training configuration, weights, and public release for YOLOv26(m). If no such release exists, recompute all rankings in Tables III–V after removing every YOLOv26(m) row and verify whether YOLOv11m becomes the top-scoring entry; if a released YOLOv26(m) exists, reproduce its VisDrone and AMVD results with the stated 640x640 input and the same training setup, and check whether the reproduced mAP values match 47.8 and 95.7. Additionally, mechanically compare the text statements in Sections VI-B and VI-C against the table values: if any table lists a score above the claimed 'highest' value, the claim is internally inconsistent.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central empirical claim is that YOLOv11m achieves the best accuracy: 46.3% mAP@0.5 on VisDrone (Section VI-B) and 95.1% on AMVD (Section VI-C). But the paper's own tables contradict this. Table III lists YOLOv26(m) at 47.8% mAP@0.5 on VisDrone; Table IV lists YOLOv26(m) at 47.8/29.4, above YOLOv11(m) at 46.3/28.8; Table V lists YOLOv26(m) at 95.7% on AMVD, above YOLOv11(m) at 95.1%. The text never names YOLOv26 outside the tables, provides no citation, no architecture description, no training configuration, and no checkpoint source. If YOLOv26(m) is a legitimate model, then the statements 'YOLOv11m achieved the highest mAP@0.5' and 'YOLOv11m architecture emerged as the superior candidate' are simply false as printed. If YOLOv26(m) is not a legitimate or standard release, then the tables contain unsupported entries, and the rankings and deployment trade-off discussion built on them are unsubstantiated. Either way, the central empirical conclusion does not follow from the presented data. The concern is load-bearing because the paper's headline findings, the model-evolution narrative, and the practical recommendation of YOLO11m or YOLO11n for UAV traffic monitoring all depend on which model actually ranks first.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript is a survey and experimental benchmarking study of vision-based vehicle detection for UAV-based traffic monitoring. It reviews architectural components of detectors and YOLO variants, describes a range of aerial datasets, and reports new comparative experiments on the VisDrone and Aerial Multi-Vehicle Detection (AMVD) datasets, together with an edge-deployment energy benchmark on an NVIDIA Jetson Orin NX. The claimed main findings are that medium YOLO variants, especially YOLOv11m, offer the best accuracy-efficiency trade-off for UAV deployment, while the nano variant YOLOv11n is the most energy-efficient choice for power-limited operations.","tokens_in":31223,"tokens_out":6654,"duration_ms":59240,"significance":"If fully supported, the paper would provide a practically useful model-selection reference for UAV traffic monitoring, combining accuracy, parameter count, latency, and energy consumption on both a public benchmark and a real-world aerial dataset. The survey component is broad and reasonably current, and the edge benchmarking across multiple power modes addresses a genuine deployment need. However, the central benchmark conclusions are not currently supportable because an undocumented model appears in the main tables and several textual claims are contradicted by those same tables; the experimental protocol also lacks key reproducibility information. The paper's significance can therefore only be assessed after the experiments are corrected, documented, and made internally consistent.","major_comments":[{"comment":"The headline claims that 'YOLOv11m achieved the highest mAP@0.5 at 46.3%' on VisDrone and that 'the YOLOv11m architecture emerged as the superior candidate' on AMVD are contradicted by the paper's own tables. Table III and Table IV report YOLOv26(m) at 47.8% mAP@0.5 on VisDrone, and Table V reports YOLOv26(m) at 95.7% on AMVD, both above the corresponding YOLOv11m values. YOLOv26(m) is never cited, described, referenced, or given a training configuration anywhere outside these tables. Please either remove the YOLOv26 entries or provide a proper citation, architecture description, and training setup, and then revise all 'highest' and 'superior' statements so that they are consistent with the final tables.","section":"Section VI-B and Section VI-C, Tables III-V"},{"comment":"The reproducibility of all benchmark results is not established. The text lists hardware and software versions but gives no learning rate, optimizer, batch size, number of epochs, augmentation pipeline, data split, or random seeds, and no table reports error bars or repeated-run statistics. This matters because several headline comparisons are close, for example YOLOv11m versus YOLOv12m on VisDrone (46.3 vs. 45.7) and on AMVD (95.1 vs. 94.6). Please report the full training protocol, and either provide variance over multiple runs or explicitly state that the numbers are single-run point estimates.","section":"Section VI-A"},{"comment":"The energy-per-frame metric is defined as E_f = P_limit/FPS, which assumes the platform draws exactly the configured power limit. On an embedded GPU with dynamic voltage and frequency scaling, actual power draw is typically below the power cap and varies with utilization, so this metric can systematically misstate J/Frame and can change efficiency rankings. Please measure or report actual power consumption, or relabel the quantity as an upper-bound proxy and discuss the limitation in the deployment analysis.","section":"Section VI-D, E_f definition"},{"comment":"The YOLOv9(c) row in Table V lists mAP@0.5 = 90.2 while the class APs are Car 97.0, Bus 92.8, and Truck 90.2; the mean of these class APs is approximately 93.3, which is inconsistent with the reported mAP. Since this table directly supports the benchmark conclusion, the entry must be corrected or the computation explained, for example if mAP is computed over a different class set or with different rounding.","section":"Table V, YOLOv9(c) row"}],"minor_comments":[{"comment":"The last two paragraphs of this subsection are duplicated: 'Real-Time Processing Constraints' is immediately followed by a second paragraph beginning 'Real-time vehicle detection is crucial for UAV applications...' with nearly identical content. One copy should be removed.","section":"Section VII-A-7"},{"comment":"The VisDrone statistics are inconsistent between the text, which reports 288 video clips, 261,908 frames, 10,209 static images, and more than 2.6 million marked boxes, and Table II, which reports '10k images, 263 videos' and '2.5M' instances. Please use consistent and clearly defined rounded values.","section":"Section V-A and Table II"},{"comment":"Notation is inconsistent: Table VI uses 'YOLO11n/s/m' while the text and other tables use 'YOLOv11...'; similarly, 'FLOPS' and 'FLOPs' are used interchangeably. Please standardize terminology throughout.","section":"Tables IV-VI and text"},{"comment":"The pNeuma description states that trajectories 'are captured along with the same frequency of 0.04-seconds computed as the highest frame rate of the video'; this sentence is unclear and should be rewritten to state the actual sampling rate.","section":"Section V-B-1"},{"comment":"The AMVD dataset was created by co-authors of this paper (reference [76]), and the evaluation protocol on it is not specified. This is not improper, but the paper should explicitly disclose the relationship and state the exact train/validation split and evaluation subset used for Table V.","section":"Section V and Section VI"},{"comment":"The sentence 'Our experiments results align with previous work on [98]–[100]' is vague and grammatically awkward; please specify which results are being compared and how the cited works support that comparison.","section":"Section VI-B"}],"recommendation":"major_revision","confidential_remarks":"The central empirical section needs a substantive revision before the paper can be considered for publication. I would ask the editor to verify whether 'YOLOv26' is a real, citable model release; if it is a typo or an unpublished variant, the tables and all conclusions derived from them must be redone. The AMVD dataset is co-authored by one of the authors, so the review process should ensure that the dataset relationship is disclosed and that the evaluation protocol is independently verifiable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the survey portion is worth a look, but the experimental core is currently self-contradictory and cannot be used as-is.\n\nWhat's actually new: the paper compiles recent (post-2022) work on UAV-based vehicle detection into a manageable overview, with a helpful comparison table of prior surveys, a reasonable taxonomy of YOLO variants, and a broad list of datasets. The Jetson Orin NX energy measurements for nano and medium models are a useful addition for deployment-focused practitioners, and the AMVD results add a new evaluation point on top-down Cyprus imagery.\n\nWhere it falls apart: the text claims YOLOv11m achieved the highest mAP on both VisDrone (46.3%) and AMVD (95.1%). But the paper's own tables list an undefined YOLOv26(m) with higher scores (47.8% and 95.7%) across Tables III, IV, and V, plus an entry in Table VI. YOLOv26 appears nowhere in the text, has no citation, no architecture description, no training configuration. That is not a minor omission—the central comparative conclusion is internally contradicted by the data presented. Either the text is wrong, or the tables contain unsupported entries. In both cases, the benchmark ranking and the deployment recommendations built on it do not follow from the paper as printed.\n\nThere are additional reproducibility gaps: no training hyperparameters, seeds, or data splits; no error bars; no code release. A duplicated paragraph in Section VII-A7 (the real-time processing bullet appears twice, nearly verbatim) points to weak editorial control. Using AMVD, a dataset co-authored by one of the authors, is a minor conflict-of-interest consideration but not a circularity issue.\n\nThe survey alone is a passable resource for someone entering the area, and the energy-per-frame data could be valuable once the model inventory is corrected. But the experimental insights—a stated contribution—are currently untrustworthy.\n\nWho this is for: practitioners or newcomers wanting a compact overview of UAV traffic monitoring datasets and YOLO variants. The experimental results, once corrected, could guide edge deployment model selection.\n\nMy recommendation: don't accept as-is, but this is fixable. Send it to review with a clear request to resolve the YOLOv26 problem, supply full experimental details, and fix the internal contradictions. A serious referee could turn this into a usable paper. If YOLOv26 cannot be sourced, the tables and all conclusions must be revised.","headline":"Useful survey, but its central experimental claim is contradicted by its own tables: an undefined YOLOv26(m) outranks YOLOv11m.","tokens_in":31784,"tokens_out":4421,"would_cite":false,"duration_ms":37912,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper's benchmark experiments claim that YOLOv11m achieves the highest vehicle detection accuracy on two aerial datasets, while the nano-scale YOLO11n consumes the least energy per frame on an embedded drone platform.","keywords":["UAV traffic monitoring","vehicle detection","YOLO","deep learning","energy efficiency","aerial benchmark","VisDrone","AMVD"],"falsifier":"Re-running the stated benchmark protocol on VisDrone2019-val with an official YOLOv26(m) checkpoint, or confirming that no such official checkpoint exists, would settle whether the reported 47.8 mAP@0.5 and the resulting ranking are reproducible.","tokens_in":30698,"feed_emoji":"🚁","tokens_out":4957,"duration_ms":42768,"temperature":0.7,"pith_summary":"This survey reviews deep-learning vehicle detection for drone-based traffic monitoring and adds its own head-to-head experiments. On the VisDrone benchmark, YOLOv11m reaches the highest mAP@0.5 at 46.3%, and on the high-altitude Cyprus dataset AMVD it reaches 95.1% mAP@0.5 with 20.0 million parameters, beating all tested variants including YOLOv12m. On an embedded edge platform, the nano-scale YOLO11n sustains 43.7 frames per second at 0.229 joules per frame under a 10-watt power limit, making it the paper's recommended choice for energy-constrained, long-endurance flights. The paper concludes that medium-sized YOLO detectors currently offer the best accuracy-efficiency trade-off for UAV traffic monitoring, while nano models are preferable when battery life is the binding constraint.","feed_headline":"YOLOv11m tops drone vehicle detection at 95.1%","feed_subtitle":"Benchmarks show medium YOLO for accuracy, nano YOLO for energy-limited drone flights.","key_machinery":"The machinery is the standardized comparative benchmark: both aerial datasets resized to 640 by 640 pixels, training on four Tesla V100 GPUs under a fixed PyTorch environment, evaluation with mAP@0.5 and mAP@0.5:0.95, and edge inference on the Jetson Orin NX across four power modes using the energy-per-frame metric $E_f = P_{\\text{limit}} / \\text{FPS}$. The central objects are the YOLO detector family spanning nano to medium scales, together with the two complementary benchmarks, VisDrone and AMVD, chosen for their contrasting altitudes and geographies.","core_discovery":"The paper's central empirical claim is a quantitative map of the accuracy-latency-energy trade-off for current YOLO detectors in UAV traffic monitoring. Trained identically at 640 by 640 pixels on VisDrone2019-val and AMVD, YOLOv11m outperforms all compared detectors, including YOLOv12m and a model listed as YOLOv26(m), while YOLO11n provides the lowest energy per frame on the Jetson Orin NX edge platform. The authors interpret this as evidence that model scale, not mere recency, drives deployable performance, and that deployment planning should be guided by an energy-per-frame metric defined as the power limit divided by the inference speed.","pith_inferences":["The results tables include a model called YOLOv26(m) with specific accuracy, parameter, and speed numbers but no description or citation anywhere in the text; this should be verified before relying on the ranking, since the paper's model-evolution discussion rests on it.","Combining the recommended YOLOv11m with slice-aided inference, which the paper cites as improving small-object average precision by up to 14.5%, would likely yield larger gains than model swaps alone.","The energy-per-frame metric could be extended to whole-system budgets that include data transmission and multi-UAV coordination, factors the paper lists as open challenges but does not quantify.","Because the two datasets are geographically narrow and mostly captured in clear daytime conditions, the rankings may not transfer to fog, rain, or night operations that the paper itself identifies as underexplored."],"forward_implications":["For precision-critical traffic tasks, a medium-scale detector such as YOLOv11m appears to be the current best choice; for battery-limited long-endurance flights, a nano-scale model such as YOLO11n is preferable.","Future evaluations of drone-deployed detectors should report energy-per-frame alongside accuracy, since frames per second alone overstates viability on power-constrained hardware.","Performance gains in recent YOLO versions concentrate on small and occluded object categories, indicating that progress in aerial detection is driven by improvements relevant to high-altitude, small-object scenarios.","The same standardized benchmark protocol can be applied to newer detector families beyond the YOLO lineage, providing a direct comparison path for future architectures."],"supporting_citations":[{"why":"Provides the VisDrone dataset, the primary benchmark for the accuracy comparisons and energy experiments.","marker":"[75]"},{"why":"Provides the Aerial Multi-Vehicle Detection Dataset used as the second, high-altitude benchmark for generalization.","marker":"[76]"},{"why":"Supplies the edge-benchmarking protocol and the energy-per-frame derivation used in the deployment analysis.","marker":"[57]"},{"why":"Introduces YOLOv12, the closest competing architecture in the comparison, whose attention-centric modules the paper contrasts with YOLOv11.","marker":"[71]"},{"why":"Defines the original YOLO one-stage detection framework that all compared variants extend.","marker":"[16]"}],"fun_headline_variants":["YOLOv11m wins drone vehicle detection, YOLO11n saves energy","YOLOv11m leads accuracy, YOLO11n leads efficiency","Drone traffic detection: YOLOv11m for precision, YOLO11n for low power","YOLOv11m beats YOLOv12m in drone vehicle detection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The benchmark rankings rest on a model called YOLOv26(m) that appears only in the results tables with no description, citation, or training details anywhere in the text, so the paper gives no way to verify its existence or that it was trained under the same conditions as the other detectors.","fun_headline_variants_meta":{"raw":{"variants":["YOLOv11m wins drone vehicle detection, YOLO11n saves energy","YOLOv11m leads accuracy, YOLO11n leads efficiency","Drone traffic detection: YOLOv11m for precision, YOLO11n for low power","YOLOv11m beats YOLOv12m in drone vehicle detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000556,"raw_usage":{"total_tokens":2615,"prompt_tokens":885,"completion_tokens":1730,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":501,"completion_tokens_details":{"reasoning_tokens":1639}},"tokens_in":501,"tokens_out":1730,"duration_ms":11361,"temperature":1.0,"reasoning_tokens":1639,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T00:32:16.437973+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-running the stated benchmark protocol on VisDrone2019-val with an official YOLOv26(m) checkpoint, or confirming that no such official checkpoint exists, would settle whether the reported 47.8 mAP@0.5 and the resulting ranking are reproducible.","supporting_citations":[],"review_version":1}