{"id":"8b5bac6c-6004-4255-8096-75031e104efd","arxiv_id":"2501.05885","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"EDNet, an edge-tuned YOLOv10 variant with an extra small-object head, Cross Concat fusion, and a Faster Context Attention block, reports up to +5.6 mAP@50 over YOLOv10 on VisDrone.","lead":"This paper introduces EDNet, a YOLOv10-based object detector for drone imagery that adds a high-resolution detection head, an attention block, a new feature-concatenation scheme, and the WIoUv3 loss. On the VisDrone benchmark it reports higher mAP than YOLOv10 baselines at several model sizes, and it measures 16 to 55 FPS on an iPhone 12 after CoreML conversion.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Baseline comparability is unverified: Table II does not establish that YOLOv5/6/8/9/10, RT-DETR, VAMYOLOX, and EdgeYOLO numbers were produced under EDNet's training protocol, so the claimed 5.6-point mAP gain and the parameter-efficiency narrative may reflect protocol differences rather than the…","rationale":"The reader's weakest assumption identifies exactly the same load-bearing concern: Table II comparisons presuppose an equivalent training and evaluation protocol across all baselines, and the paper provides no evidence of that equivalence. I do not see an independent objection that would push the verdict to REJECT, because the paper contains no internal mathematical contradiction and the proposed architectural components are plausible. However, the central claim of consistent outperformance and parameter efficiency is only as strong as the comparability of the baselines. Since the reader already assigned CONDITIONAL with moderate confidence, my stress-test does not change the verdict; it reinforces the condition that baseline provenance and identical-protocol reproduction be supplied. A concrete retraining check would settle the concern: if EDNet retains its lead over same-size YOLOv10 variants under the paper's own stated recipe, the headline claim survives; if the gap collapses, the paper's central contribution is mostly protocol-dependent. The additional observation about same-size parameter counts is a secondary caveat that further supports the CONDITIONAL verdict but does not by itself invalidate the accuracy comparison, since higher parameters can still yield better accuracy at the reported latencies.","tokens_in":11322,"tokens_out":4043,"duration_ms":40923,"concrete_test":"Retrain the full set of YOLOv10 variants (N, S, M, B, L, X) and at least one external baseline such as YOLOv8-M from scratch using exactly the Section III-B protocol (SGD, lr=0.01, momentum=0.9, 200 epochs) on the same VisDrone training split, then evaluate with the same script; report mean and standard deviation over at least three seeds. If EDNet's same-size advantage over YOLOv10 shrinks below the reported 5.6 mAP@50 points or falls within baseline run-to-run variance, the central comparative claim is not established as stated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central accuracy claim is that EDNet 'consistently outperforms its competitors across all model sizes' (abstract; Section III-C). This claim rests entirely on Table II's cross-model comparisons, but the paper never states whether baseline numbers were retrained in-house with the protocol in Section III-B (SGD, lr=0.01, momentum=0.9, 200 epochs, VisDrone train) or copied from the respective source papers. This is load-bearing because training schedules, input resolution, augmentation, and validation procedure all move mAP by several points on VisDrone. Notably, the ablation 'Baseline' in Table III (20.4M params, 43.4 mAP@50) matches the YOLOv10-B entry in Table II, which suggests at least one baseline was trained by the authors, but no equivalent statement covers YOLOv5/6/8/9, RT-DETR, VAMYOLOX, or EdgeYOLO. If those numbers come from different protocols, the 5.6-point gap over YOLOv10-M and the 'superior efficiency' narrative could be partially artifactual. The concern is amplified by the absence of error bars (all results appear to be single runs) and by the fact that the validation split is used both for checkpoint selection and for final reporting. There is also a secondary internal tension in the efficiency claim: at matched sizes EDNet has more parameters than YOLOv10 (EDNet-S 9.3 vs 8.0, EDNet-M 19.1 vs 16.5, EDNet-L 31.7 vs 25.7, EDNet-X 48.7 vs 31.6), so the 'fewer parameters' statement depends on comparisons against larger models rather than same-size ones. These issues are addressable with better reporting, but they are the weakest load-bearing support for the headline comparison.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes EDNet, a YOLOv10-based detector for small objects in UAV imagery. The main architectural contributions are a C2f-FCA block with a Faster Context Attention mechanism, an additional XSmall detection head, a Cross Concat feature-fusion strategy, and the WIoUv3 bounding-box loss. The authors report seven model sizes, evaluate on VisDrone validation, and claim consistent improvements over YOLOv5/6/8/9/10 and RT-DETR, with up to a 5.6 mAP@50 gain over YOLOv10-M, plus real-time deployment measurements on iPhone, iPad, Mac, and Raspberry Pi. The paper also includes an ablation study and hardware acceleration via CoreML, OpenVINO, TensorRT, and ONNX.","tokens_in":11610,"tokens_out":4949,"duration_ms":47513,"significance":"If the empirical claims hold, the paper would provide a useful accuracy-efficiency trade-off for edge-based UAV detection, with a practical deployment story across multiple hardware platforms. The authors release code and pretrained models, report seven scalable variants, and provide per-component ablations. These are concrete strengths. However, the central comparative claim rests on cross-model comparisons whose training and evaluation protocols are not verified, and the reported numbers lack error bars and suffer from validation-set selection bias. The contribution is therefore plausible but not yet rigorously established.","major_comments":[{"comment":"The paper never states whether the YOLOv5/6/8/9/10, RT-DETR, VAMYOLOX, and EdgeYOLO numbers in Table II were retrained in-house under the protocol described in Section III-B or copied from their source papers. mAP on VisDrone is sensitive to training schedule, input resolution, augmentation, and validation procedure, so the claimed 5.6-point gain over YOLOv10-M and the 'significantly fewer parameters' narrative could partly reflect protocol differences rather than architectural improvements. Please provide per-row provenance, and ideally retrain all baselines under the same protocol; at minimum, clearly mark literature-derived values and discuss the sensitivity of the comparison to protocol differences.","section":"Section III-B and Table II"},{"comment":"All results appear to be single runs with no error bars or multiple seeds. The differences in the N and S tiers are small enough that run-to-run variation could change the ranking (e.g., EDNet-N 33.8 vs YOLOv10-N 32.3; EDNet-S 42.5 vs YOLOv8-S 38.5). Please report mean and standard deviation over at least three independent training runs, or otherwise provide evidence that the observed gaps are statistically meaningful.","section":"Section III-B and Table II"},{"comment":"Section III-B states that 'the best-performing model is selected during the training process,' and Table II reports results on the VisDrone validation split. This means the validation set is used both for checkpoint selection and for the headline comparisons, which introduces selection bias and can overstate generalization performance. Please use a held-out test split for final comparisons, or otherwise account for the selection procedure when reporting validation numbers.","section":"Section III-B and III-C"},{"comment":"The claim of 'significantly fewer parameters' is only true when comparing EDNet to larger YOLOv10 variants. At equal nominal sizes, EDNet has more parameters than YOLOv10 in Table II (e.g., EDNet-S 9.3M vs YOLOv10-S 8.0M; EDNet-M 19.1M vs YOLOv10-M 16.5M; EDNet-L 31.7M vs YOLOv10-L 25.7M; EDNet-X 48.7M vs YOLOv10-X 31.6M). The efficiency narrative should be rephrased to make these cross-size comparisons explicit, or supplemented with a same-parameter comparison.","section":"Section III-C and Table II"},{"comment":"The text makes quantitative comparisons to VAMYOLOX-X (47.6% mAP@50, 104.6M parameters) and EdgeYOLO-Tiny (5.5M parameters), yet Table II explicitly excludes these models 'to ensure a fair comparison' because of hardware/software differences. This is internally inconsistent: either include them with clearly stated provenance and protocol, or avoid drawing quantitative superiority conclusions from literature numbers obtained under different training and evaluation conditions.","section":"Section III-C"}],"minor_comments":[{"comment":"The CAA attention equation computes the weighting matrix A, but the manuscript does not show how A is applied to the features before the residual addition; please add the missing multiplication or output definition so the block is fully specified.","section":"Section II-A, Eqs. (1)-(3)"},{"comment":"Equation (7) is typeset ambiguously: the formula for r and the roles of alpha and delta are unclear as printed. Please rewrite it in a standard one-line form.","section":"Section II-C, Eq. (7)"},{"comment":"The text says the dataset has 10 target categories but lists only nine: pedestrian, people, bicycle, car, van, tricycle, awning-tricycle, bus, and motorbike. The missing category (truck) should be added.","section":"Section III-A"},{"comment":"The ablation table would be easier to interpret if it also reported the change in FLOPs or latency for each incremental component, since the text discusses computational overhead but the table only lists parameter counts.","section":"Table III"},{"comment":"The sentence 'all sizes of EDNet’s efficiency is proved on iPhone' is grammatically awkward and overstates what a latency measurement can prove; please rephrase to 'the efficiency of all EDNet sizes was demonstrated on an iPhone.'","section":"Section IV"}],"recommendation":"major_revision","confidential_remarks":"The core architectural work is reasonable and the deployment measurements are a useful asset. However, the central comparative claim cannot be verified as presented because baseline provenance is unstated, no repeated runs are reported, and validation-based selection is used for the headline numbers. These issues are fixable within the scope of the manuscript by retraining baselines or clearly labeling literature values, adding error bars, and using a held-out test split. I would recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: EDNet is a well-executed engineering paper. It assembles known parts—a P2-style high-resolution head, a cross-concatenation in the neck, a FasterNet/PKINet-style attention block, and WIoUv3—into a YOLOv10-based detector family for UAV imagery, and reports solid gains on VisDrone (up to +5.6 mAP@50 over YOLOv10) with a 1.78M-parameter Tiny variant and 16–55 FPS on an iPhone 12. The code and pretrained models are public, which is real evidence and makes the work reproducible.\n\nWhat's genuinely useful: the seven-size family covers a deployment spectrum from Raspberry Pi to Jetson-class hardware, the hardware acceleration results (CoreML, OpenVINO, TensorRT, ONNX) are concrete, and the ablation shows a coherent story—each addition (WIoU, XSmall head, Cross Concat, C2f-FCA) improves mAP while the FCA block reduces parameters. The XSmall head and Cross Concat are simple and easy to adopt in other detectors.\n\nThe soft spots are real but not fatal. The main one: Table II never says whether baseline numbers (YOLOv5/6/8/9, RT-DETR, VAMYOLOX, EdgeYOLO) were retrained under the same protocol or taken from their papers. Training schedules and augmentation move VisDrone mAP by several points, so the 5.6-point gap could be partly an artifact of protocol differences. The ablation's baseline matches YOLOv10-B in parameters, suggesting at least that one was trained in-house, but the same is not stated for the others. There are no error bars or multi-seed results, and model selection and final evaluation share the same validation split. The parameter-efficiency narrative also cherry-picks comparisons against larger models; at matched sizes EDNet has more parameters than YOLOv10, though it still wins on mAP.\n\nThese are reporting issues rather than load-bearing contradictions. The paper's logic is coherent on its own terms, and the public code means a reader can verify the central claim. I'd send this to peer review. A referee should ask for baseline provenance, multiple seeds, and a clearer statement on validation-based model selection. The engineering contribution is worth the referee's time.","headline":"EDNet is a competent engineering paper whose headline gains over YOLOv10 are plausible but rest on a comparison table with unverified baseline protocols; still worth refereeing.","tokens_in":12266,"tokens_out":2661,"would_cite":true,"duration_ms":24341,"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":"EDNet, a YOLOv10-based detector with an extra high-resolution head, cheaper context attention, and a new fusion strategy, claims up to 5.6 mAP@50 gain over YOLOv10 on VisDrone while running in real time on edge devices.","keywords":["small object detection","UAV imagery","YOLOv10","edge computing","feature fusion","attention mechanism","VisDrone","real-time inference"],"falsifier":"Retrain YOLOv10-M under EDNet's exact protocol (200 epochs, SGD lr 0.01, momentum 0.9, same input resolution and augmentation) and evaluate on the VisDrone validation split; if it reaches or exceeds 47.1% mAP@50, the architecture would not be doing the work the comparison attributes to it.","tokens_in":11032,"feed_emoji":"🚁","tokens_out":7161,"duration_ms":60299,"temperature":0.7,"pith_summary":"The paper proposes EDNet, a detection framework built on YOLOv10, and argues that four modifications—an extra high-resolution detection head, a cross-concatenation feature fusion strategy, a C2f-FCA block with faster context attention, and the WIoUv3 box loss—let one architecture beat all compared state-of-the-art detectors on VisDrone while shrinking parameter counts. The central claim is that EDNet consistently outperforms competitors at every model size, with up to a 5.6 mAP@50 gain over YOLOv10 and a 33.3% mAP@50 for a 1.78M-parameter Tiny variant. The motivation is practical: drone imagery is dominated by tiny targets, and edge devices need low-latency, low-memory inference without post-processing. If the claim holds, EDNet offers a scalable accuracy-efficiency trade-off for real-time UAV detection on phones and embedded boards.","feed_headline":"EDNet tops YOLOv10 by 5.6 mAP on drone imagery","feed_subtitle":"A YOLOv10 variant with an extra small-target head and cheaper attention runs 16-55 FPS on an iPhone 12.","key_machinery":"The load-bearing pieces are four. C2f-FCA is a YOLOv10 C2f variant whose bottleneck blends a partial convolution (from FasterNet) with two pointwise convolutions and a Context Anchor Attention sub-block that uses average pooling, a pointwise convolution, and two depth-wise strip convolutions to approximate large-kernel attention cheaply; this is what the paper credits for better feature extraction and parameter reduction. The XSmall detection head adds a 160x160 feature map (downsampled only twice) so tiny objects retain more pixels. The Cross Concat strategy feeds the SPPF output into the cross-concatenation while keeping the PSA output connected to the first upsampling block, giving the neck multi-scale context. WIoUv3 is a dynamic focusing box-loss that down-weights outlier samples via a non-monotonic focusing coefficient. Together these carry the claimed accuracy gain; the ablation attributes the largest single jump (4.2 mAP@50) to the XSmall head plus CCS, and the final C2f-FCA step both improves accuracy and cuts parameters from 29.9M to 25.5M.","core_discovery":"EDNet's core discovery is that a YOLOv10 baseline can be made substantially better at small-target detection in aerial images by adding an XSmall detection head at 160x160 resolution, rerouting the SPPF output through a Cross Concat strategy in the neck, replacing the backbone's C2f blocks with C2f-FCA blocks that combine partial convolution and Context Anchor Attention, and switching the box regression loss to WIoUv3. On the VisDrone 2019-DET-val set, EDNet-M reaches 47.1% mAP@50 versus 41.5% for YOLOv10-M, and EDNet-X reaches 50.2% versus 44.8% for YOLOv10-X, with fewer parameters in several comparisons. The paper attributes the gains mainly to richer small-object features from the extra head and the multi-scale context supplied by the cross-concatenation, while the FCA block reduces parameters and adds speed. It also reports real-time deployment: all seven EDNet variants run at 16 to 55 FPS on an iPhone 12 after INT8/FP16 optimization and neural-engine acceleration.","pith_inferences":["If the gains are reproducible under identical training protocols, the same combination—extra high-resolution head, cross-concatenation, partial-convolution attention, and wise-IoU loss—could transfer to other YOLO-era detectors or to remote-sensing datasets with tiny objects, though the paper only tests VisDrone.","The XSmall head adds a 160x160 detection map, which raises compute; the paper's numbers suggest C2f-FCA compensates, but on lower-end CPUs the net speed benefit may shrink—a testable extension would measure per-component latency on ARM CPUs rather than only on A100 and iPhone.","Because the FCA block's strip convolutions approximate large-kernel attention at low cost, the design hints at a general low-parameter route to context modeling for small objects, but the paper does not isolate whether the gain comes from the attention mechanism or from the partial-convolution structure."],"forward_implications":["At every size tier in Table II, the EDNet variant beats the corresponding YOLOv3/5/6/8/9/10, RT-DETR, VAMYOLOX, and EdgeYOLO numbers on VisDrone val, so the design recipe is claimed to transfer across model capacities.","A 1.78M-parameter Tiny variant reaches 33.3% mAP@50, surpassing all compared N-size YOLO models, which would make Raspberry Pi-class edge targets viable.","Because YOLOv10 needs no NMS post-processing, EDNet inherits end-to-end inference; on iPhone 12 the seven variants span 16 to 55 FPS, and on a Raspberry Pi 5 the optimized Tiny model is 3.1 times faster than the raw implementation.","The ablation suggests each component contributes: WIoU adds 0.3, the XSmall head adds 3.7, CCS adds 0.5, and C2f-FCA adds 0.4 while removing 4.4M parameters, implying the efficiency story rests mainly on the FCA block."],"supporting_citations":[{"why":"YOLOv10 is the baseline detector EDNet extends; its NMS-free design and architecture define the starting point and the direct comparison.","marker":"[6]"},{"why":"VisDrone is the dataset used for all training and validation; every reported mAP number is measured on its validation split.","marker":"[1]"},{"why":"WIoUv3 supplies the box-regression loss that the paper credits for handling noisy and outlier samples.","marker":"[27]"},{"why":"FasterNet provides the partial-convolution design at the core of the C2f-FCA block.","marker":"[25]"},{"why":"Poly Kernel Inception Network contributes the Context Anchor Attention sub-block used in the FCA bottleneck.","marker":"[26]"},{"why":"RT-DETR is the transformer-based real-time detector compared in the main results.","marker":"[28]"},{"why":"EdgeYOLO is the edge-optimized detector used as the resource-efficiency baseline for the comparison.","marker":"[18]"},{"why":"YOLOv5 provides a major YOLO baseline and the ConvBNSiLU blocks that carry over into EDNet.","marker":"[20]"}],"fun_headline_variants":["EDNet beats YOLOv10 by 5.6 mAP on drone imagery","Small targets, big gain: EDNet ups YOLOv10 by 5.6 mAP","EDNet: extra head, cheaper attention, 5.6 mAP gain over YOLOv10","EDNet runs 16-55 FPS on iPhone with 5.6 mAP boost over YOLOv10"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central comparison assumes that all baseline models in Table II were trained and evaluated under the same protocol as EDNet (epochs, optimizer, resolution, augmentation, validation), so if any baseline numbers come from different settings, the claimed gains could be artifacts of protocol rather than architecture.","fun_headline_variants_meta":{"raw":{"variants":["EDNet beats YOLOv10 by 5.6 mAP on drone imagery","Small targets, big gain: EDNet ups YOLOv10 by 5.6 mAP","EDNet: extra head, cheaper attention, 5.6 mAP gain over YOLOv10","EDNet runs 16-55 FPS on iPhone with 5.6 mAP boost over YOLOv10"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00075,"raw_usage":{"total_tokens":3370,"prompt_tokens":1005,"completion_tokens":2365,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":621,"completion_tokens_details":{"reasoning_tokens":2260}},"tokens_in":621,"tokens_out":2365,"duration_ms":15287,"temperature":1.0,"reasoning_tokens":2260,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:06:15.797166+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain YOLOv10-M under EDNet's exact protocol (200 epochs, SGD lr 0.01, momentum 0.9, same input resolution and augmentation) and evaluate on the VisDrone validation split; if it reaches or exceeds 47.1% mAP@50, the architecture would not be doing the work the comparison attributes to it.","supporting_citations":[{"cited_title":"Zhu et al., ”Detection and Tracking Meet Drones Challenge,” in IEEE Transactions on Pattern Analysis and Machine Intel- ligence, vol","cited_arxiv_id":null,"evidence_quote":"VisDrone is the dataset used for all training and validation; every reported mAP number is measured on its validation split."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"FasterNet provides the partial-convolution design at the core of the C2f-FCA block."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Poly Kernel Inception Network contributes the Context Anchor Attention sub-block used in the FCA bottleneck."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"RT-DETR is the transformer-based real-time detector compared in the main results."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"EdgeYOLO is the edge-optimized detector used as the resource-efficiency baseline for the comparison."}],"review_version":1}