{"id":"ed79de67-08e0-4771-bf2e-558337267af8","arxiv_id":"2507.12727","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A YOLOv8 variant using ASF feature fusion, a P2 detection head, and Soft-NMS reports 36.1% higher mAP50:95 on VisDrone2019-DET than its YOLOv8-m baseline.","lead":"This paper combines three existing techniques, a scale fusion module, a high-resolution detection layer, and soft non-maximum suppression, to make a YOLOv8 drone detector better at spotting tiny objects. On the VisDrone benchmark it reports a large jump in average precision over the baseline, but the results come from a single run without error bars.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Claimed improvement is dominated by a suspiciously large Soft-NMS gain and an inconsistent final mAP, leaving the central result unverified.","rationale":"The reader identified the fairness and single-run nature of the baseline as the weakest assumption, which is legitimate but broad. My stress test sharpens this into a specific, testable concern: the large Soft-NMS gain is the single most load-bearing element of the empirical claim, because it contributes the majority of the reported improvement and is not supported by typical Soft-NMS behavior or by any reported NMS settings. The 0.351 vs 0.352 discrepancy further indicates that the evaluation pipeline is not fully consistent. I agree with the reader's CONDITIONAL verdict: the paper needs independent replication and error bars, plus a same-weights NMS comparison, before the central claims of a 36.1% improvement and the contribution of each component can be accepted. The concern does not currently move the verdict because it is plausible that the gain could reproduce with the released code and proper NMS tuning; it simply needs the concrete checks I propose. Thus UNCHANGED, with the condition that these checks be performed.","tokens_in":8503,"tokens_out":3044,"duration_ms":34097,"concrete_test":"Using the released repository, load the published baseline YOLOv8-m weights and the final SOD-YOLO weights. Run both on VisDrone2019-DET-val under identical settings, varying only the NMS type: standard NMS and Soft-NMS, with the same confidence threshold and IoU threshold as in the repo's default evaluation. If the mAP50:95 difference between NMS and Soft-NMS is below 0.01, the Soft-NMS claim fails. Also train the +ASF+P2 variant (without Soft-NMS) using the same hyperparameters for at least 3 random seeds, and compare its mAP50:95 to the published 0.294; if the standard deviation exceeds 0.005, the ablation increments are not statistically distinguishable.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests on the ablation in Table 2, where switching from NMS to Soft-NMS as post-processing changes mAP50:95 from 0.294 to 0.352 and mAP50 from 0.476 to 0.526, purely at inference time (FLOPs stay 94.9G). This Soft-NMS contribution of +0.058 mAP50:95 and +0.050 mAP50 accounts for roughly 62% of the total reported gain over the YOLOv8-m baseline. Such a large gain from re-scoring overlapping boxes is atypical; Soft-NMS typically yields 1-2 mAP points on standard benchmarks. The paper does not report the NMS confidence threshold or IoU threshold used for either baseline or ablation, nor does it provide a same-weights NMS-vs-Soft-NMS comparison. If the baseline was evaluated with a poorly tuned NMS, the whole improvement hierarchy is an artifact. Additionally, the final SOD-YOLO is 0.351 mAP50:95 in Table 1 but 0.352 in Table 2, an internal inconsistency of 0.001 that matters when the headline claims a precise 36.1% relative gain. The ASF contribution is also only +0.007, so the model gains are concentrated in the post-processing step rather than in the architectural changes described as the paper's contributions.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript proposes SOD-YOLO, a modification of YOLOv8-m for small object detection in UAV imagery. The model adds an Attentional Scale Sequence Fusion (ASF) module and a P2 detection head to the neck, and replaces NMS with Soft-NMS at inference. On the VisDrone2019-DET validation set the authors report mAP50:95 of 0.351 and mAP50 of 0.526, versus 0.258 and 0.436 for the YOLOv8-m baseline, and an ablation table attributing gains to ASF (+0.007), P2 (+0.036), and Soft-NMS (+0.058) in mAP50:95. The authors state that code, hyperparameters, and model weights are publicly available.","tokens_in":8714,"tokens_out":5290,"duration_ms":52743,"significance":"If the reported gains are reproducible, SOD-YOLO would be a practically useful detector for small objects in aerial imagery, with a favorable accuracy/complexity trade-off compared with the YOLO variants in Table 1. The paper is honest in reporting a monotonic ablation and in releasing code and weights. However, the empirical evidence as presented is not yet sufficient to establish the central claim: the results come from a single run, the post-processing comparison is underspecified, the dominant gain comes from Soft-NMS rather than the architectural contributions, and there is an inconsistency in the reported final mAP. These are fixable with additional experiments and reporting, rather than flaws in the method itself.","major_comments":[{"comment":"The Soft-NMS gain of +0.058 mAP50:95 and +0.050 mAP50 (from 0.294/0.476 to 0.352/0.526) amounts to roughly 62% of the total improvement over baseline, which is much larger than the 1–2 mAP-point gains typically reported for Soft-NMS on standard benchmarks. The paper does not report the NMS/Soft-NMS IoU threshold, the confidence threshold, or a same-weights comparison in which only the post-processing rule is changed. Without those details, it is possible that the baseline NMS was run with a poorly tuned threshold, which would inflate the entire improvement hierarchy. Please provide the thresholds used, perform an ablation over NMS and Soft-NMS thresholds on identical detector outputs, and report precision-recall curves for the affected settings.","section":"Table 2, §4.6"},{"comment":"The final SOD-YOLO model is reported as 0.351 mAP50:95 in Table 1 and in Section 4.5, but as 0.352 mAP50:95 in Table 2 and Section 4.6. The headline result is a precise 36.1% relative gain (0.093 over 0.258), so an unexplained 0.001 discrepancy in the third decimal is material to the paper's central quantitative claim. Please reconcile the numbers and state which configuration produced the final model.","section":"Tables 1 and 2"},{"comment":"All experiments appear to be based on a single training run, with no error bars, confidence intervals, or significance tests reported in Tables 1 and 2. This is a load-bearing issue because the ASF increment (+0.007 mAP50:95) and some inter-model gaps are small relative to typical run-to-run variation in detection training. Please run at least three seeds per configuration, report mean and standard deviation, and indicate whether the differences between SOD-YOLO and the baseline are statistically significant.","section":"§4.3"},{"comment":"The claim that 'all models are trained and evaluated under the same settings' is not verifiable from the manuscript. The only hyperparameters listed are for SOD-YOLO; no augmentation schedule, image preprocessing, NMS configuration, or training recipe is given for the comparison models, and it is not stated whether YOLOv5-m, YOLOv7-m, YOLOv9-gelan-c, and YOLOv10-l were retrained from the same data pipeline or taken from external reports. Please specify the exact training protocol for every model in Table 1, or restrict the comparison to models trained in-house under identical conditions.","section":"§4.3, Table 1"}],"minor_comments":[{"comment":"The text states that SOD-YOLO is compared with Edge-YOLO, but Edge-YOLO does not appear in Table 1 and no quantitative Edge-YOLO result is given.","section":"§4.5"},{"comment":"Reference [19], cited for SPPF, is a document titled 'Session Peering Provisioning Framework (SPPF)' and is unrelated to the SPPF layer used in YOLO; this citation should be replaced with the appropriate source. Reference [10], cited for YOLOv5-m, points to the original YOLO paper rather than to YOLOv5.","section":"References"},{"comment":"The architecture diagram uses several abbreviations ('CZC', 'Zoom_Cat', 'Concat3C') and labels ('ASF-Attention ScalSeq') that are not defined in the text or caption, making the diagram difficult to interpret.","section":"Figure 1"},{"comment":"The description of the ASF 'attention model' is underspecified: it does not state the channel reduction ratio, the kernel sizes of the local attention, or how the two input feature maps are selected, which prevents reproduction of the neck from the text alone.","section":"§3.1"},{"comment":"There are numerous typographical and formatting issues, including inconsistent spacing in 'UA V', a missing space after a period in Section 4.5 ('baseline.Additionally'), and inconsistent notation for the NMS threshold (N_t versus Nt).","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"Given that the authors state the code and weights are publicly available, the editor may wish to have the repository checked during revision, especially the NMS configuration and the exact training settings. The discrepancy between Table 1 and Table 2 and the unusually large Soft-NMS gain are the main concerns. The paper fits the journal's scope if the empirical claims are verified."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nShort version: the paper is a competent YOLOv8-m engineering exercise for small-object detection in UAV images, with an ablation that shows monotonic gains from each added component. The central claim is a large jump on VisDrone2019-DET (0.351 vs. 0.258 mAP50:95), and they actually ship code, hyperparameters, and weights. If you work on practical drone detection, this is worth a look.\n\nWhat is new is narrow: combining an ASF-style scale-sequence neck, a P2 head, and Soft-NMS in one YOLOv8-m variant. Each ingredient is known; the novelty is the specific integration and the reported benchmark numbers. The ablation is honest in direction, and the authors acknowledge the real-time cost of their additions, which gets credit.\n\nSoft spots, in rough order of seriousness. First, the Soft-NMS contribution in the ablation is +0.058 mAP50:95 and +0.050 mAP50, which is roughly 62% of the total gain over baseline and looks atypical for a pure post-processing change. The paper does not state the NMS/Soft-NMS confidence or IoU thresholds, nor does it give a same-weights NMS-vs-Soft-NMS comparison. If the baseline NMS was poorly tuned, the whole hierarchy is an artifact. Second, the same final model is 0.351 in Table 1 and 0.352 in Table 2; a 0.001 discrepancy is small but sloppy when the headline is a precise 36.1% relative gain. Third, everything is single-run, no error bars or significance tests, so the real uncertainty is unknown. Fourth, the ScalSeq module is presented as “our proposed” when the paper’s own cited ASF-YOLO already describes the same scale-sequence fusion; that needs a correction. Fifth, the reference list has errors: [10] points to YOLOv1 instead of YOLOv5-m, and [19] is the Session Peering Provisioning Framework rather than a YOLOv8 SPPF source.\n\nNone of these are fatal in themselves, but together they make the headline improvement unverified. The central claim may well be true; the reporting just does not support the precision of the claim.\n\nThis paper deserves a serious referee, not a desk reject. A good reviewer could quickly test the Soft-NMS sensitivity and ask for multi-run numbers. I would send it to review, expecting heavy revision.\n\nBest,\n[You]","headline":"A plausible YOLOv8 integration for small-object detection on VisDrone, but the headline gain leans heavily on a suspiciously large Soft-NMS effect and the empirical reporting needs tightening.","tokens_in":9264,"tokens_out":1754,"would_cite":false,"duration_ms":21842,"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":"SOD-YOLO, a YOLOv8-m variant with scale-sequence fusion, a P2 head, and Soft-NMS, raises VisDrone mAP50:95 from 0.258 to 0.351.","keywords":["small object detection","UAV imagery","YOLOv8","attentional scale sequence fusion","P2 detection head","Soft-NMS","VisDrone2019-DET"],"falsifier":"Run both YOLOv8-m and SOD-YOLO from scratch with five different random seeds under the exact same training recipe, and re-tune the baseline's NMS and confidence thresholds; if the mean mAP$_{50:95}$ gap falls below about 0.05 or the baseline averages above 0.27, the paper's central claim of a large practical gain would be weakened.","tokens_in":8270,"feed_emoji":"🎯","tokens_out":17290,"duration_ms":146695,"temperature":0.7,"pith_summary":"Small object detection remains difficult for one-stage detectors like YOLO, especially in drone imagery where targets often occupy very few pixels. This paper proposes SOD-YOLO, a YOLOv8-m variant that adds three components: an attention-based multi-scale feature fusion (ASF) in the neck, a dedicated small-object detection head on high-resolution P2 features, and Soft-NMS post-processing. On the VisDrone2019-DET validation set, the authors report that SOD-YOLO raises mAP$_{50:95}$ from 0.258 to 0.351 and mAP$_{50}$ from 0.436 to 0.526, a 36.1% and 20.6% relative gain over YOLOv8-m. An ablation shows the three components contribute cumulatively rather than redundantly. The intended contribution is a practical, efficient recipe for improving small-object detection in UAV imagery without replacing the backbone.","feed_headline":"SOD-YOLO beats YOLOv8 by 36% on drone small-object detection","feed_subtitle":"A neck fusion, a high-res head, and Soft-NMS lift VisDrone mAP50:95 from 0.258 to 0.351.","key_machinery":"The mechanism is the combination of three targeted additions to YOLOv8-m. (1) Scale Sequence (ScalSeq) fusion: three neck features P3, P4, P5 are convolved with $1\\times1$ kernels to unify channels, upsampled to P3's resolution, stacked along a 'scale' dimension, and merged by a $1\\times1\\times1$ 3D convolution, batch norm, LeakyReLU, and 3D max pooling, replacing plain concatenation in the neck; an attention block (channel attention plus local attention) then refines the fused feature. (2) A P2 detection head: a high-resolution feature map from an early backbone stage is upsampled, concatenated with shallow features, refined by a C2f block and a ScalSeq module, and added to the P3–P5 heads, preserving fine spatial detail for tiny objects. (3) Soft-NMS: instead of zeroing overlapping detections, the post-processor decays their confidence by a factor of $1-\\mathrm{IoU}$, retaining true positives in dense clusters. The paper credits the whole improvement to the joint effect of these components, with Soft-NMS as the single largest contributor to the final total.","core_discovery":"The paper's central claim is that a YOLOv8-m detector augmented with an attentional scale-sequence fusion neck (ASF), a high-resolution P2 detection head, and Soft-NMS post-processing—collectively called SOD-YOLO—materially improves small-object detection in UAV imagery. The reported results on VisDrone2019-DET-val place SOD-YOLO at 0.351 mAP$_{50:95}$ and 0.526 mAP$_{50}$, gains of 0.093 and 0.090 over the YOLOv8-m baseline (0.258 and 0.436), which the authors frame as a 36.1% and 20.6% relative improvement. An ablation study shows each ingredient contributes: the ASF neck adds +0.007 mAP$_{50:95}$, adding the P2 head yields +0.036, and adding Soft-NMS yields a final cumulative +0.094 over baseline, with FLOPs rising from 78.7G to 94.9G and parameters at 22.6M, below the baseline's 25.8M. The comparison in Table 1 also places SOD-YOLO above YOLOv9-gelan-c and YOLOv10-l on both metrics.","pith_inferences":["The large final gain attributed to the full combination (+0.094 mAP$_{50:95}$) may partly reflect an untuned baseline: a grid search over NMS and confidence thresholds for YOLOv8-m could narrow the gap, since Soft-NMS alone usually yields far smaller gains in other detectors.","Because over 75% of VisDrone objects occupy under 0.1% of the image area, the P2 head plus scale-sequence fusion recipe is a plausible transferable template for other tiny-object domains such as satellite or medical imaging, though the paper only evaluates on VisDrone.","The reported final mAP$_{50:95}$ differs slightly between Table 1 (0.351) and Table 2 (0.352); reconciling this and reporting multi-seed means would strengthen confidence that the 0.093-point gain is not noise.","The ablation's Soft-NMS contribution could be tested directly: run the +ASF+P2 model with hard NMS at several IoU thresholds and with Soft-NMS; the difference would isolate the post-processing contribution more cleanly than the current cumulative row."],"forward_implications":["If the reported numbers hold, a YOLOv8-m user can raise VisDrone-class small-object mAP$_{50:95}$ by roughly a third using only neck fusion, a P2 head, and post-processing, without changing the backbone or adding heavy parameters.","On VisDrone2019-DET-val, SOD-YOLO's 0.351 mAP$_{50:95}$ and 0.526 mAP$_{50}$ exceed those of YOLOv9-gelan-c (0.305 and 0.489) and YOLOv10-l (0.286 and 0.462), with 22.6M parameters and 94.9G FLOPs.","The ablation's monotonic gains imply the three components are complementary: ASF improves feature fusion, P2 adds spatial resolution, and Soft-NMS recovers true positives lost by hard NMS.","The authors state that source code, hyper-parameters, and model weights are released, making the exact training recipe reproducible on a single GPU."],"supporting_citations":[{"why":"Provides the VisDrone2019-DET dataset and the small-object statistics (over 75% of objects under 0.1% of image area) used for all comparisons.","marker":"[14]"},{"why":"Supplies the attentional scale-sequence fusion (ScalSeq/ASF) mechanism that the paper adapts into the neck.","marker":"[15]"},{"why":"Defines the confidence-decay suppression rule that replaces hard NMS in SOD-YOLO's post-processing.","marker":"[18]"},{"why":"Serves as a comparison baseline in Table 1 for detector performance on VisDrone2019-DET-val.","marker":"[10]"},{"why":"A state-of-the-art baseline that SOD-YOLO must beat on the same validation set.","marker":"[12]"},{"why":"Another recent one-stage detector used as a comparison in Table 1.","marker":"[13]"}],"fun_headline_variants":["SOD-YOLO adds 36% to drone small-object mAP","Small UAV targets: +36% mAP with SOD-YOLO","UAV small-object detection up 36% via SOD-YOLO","YOLOv8 gets 36% boost for tiny drone objects","SOD-YOLO: 36% better on VisDrone small objects"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported gains rest on the assumption that the YOLOv8-m baseline's single-run scores (0.258 mAP$_{50:95}$ and 0.436 mAP$_{50}$) are representative and that all models in Table 1 were trained under identical settings; if the baseline is even slightly low, the headline 36.1% relative improvement is overstated.","fun_headline_variants_meta":{"raw":{"variants":["SOD-YOLO adds 36% to drone small-object mAP","Small UAV targets: +36% mAP with SOD-YOLO","UAV small-object detection up 36% via SOD-YOLO","YOLOv8 gets 36% boost for tiny drone objects","SOD-YOLO: 36% better on VisDrone small objects"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000268,"raw_usage":{"total_tokens":1645,"prompt_tokens":1002,"completion_tokens":643,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":618,"completion_tokens_details":{"reasoning_tokens":545}},"tokens_in":618,"tokens_out":643,"duration_ms":6940,"temperature":1.0,"reasoning_tokens":545,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T16:39:59.138541+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run both YOLOv8-m and SOD-YOLO from scratch with five different random seeds under the exact same training recipe, and re-tune the baseline's NMS and confidence thresholds; if the mean mAP$_{50:95}$ gap falls below about 0.05 or the baseline averages above 0.27, the paper's central claim of a large practical gain would be weakened.","supporting_citations":[{"cited_title":"VisDrone-DET2019: The vision meets drone object detection in im- age challenge results,","cited_arxiv_id":null,"evidence_quote":"Provides the VisDrone2019-DET dataset and the small-object statistics (over 75% of objects under 0.1% of image area) used for all comparisons."},{"cited_title":"ASF-YOLO: A novel YOLO model with attentional scale sequence fusion for cell instance segmentation,","cited_arxiv_id":null,"evidence_quote":"Supplies the attentional scale-sequence fusion (ScalSeq/ASF) mechanism that the paper adapts into the neck."},{"cited_title":"Soft-NMS: improving object detection with one line of code,","cited_arxiv_id":null,"evidence_quote":"Defines the confidence-decay suppression rule that replaces hard NMS in SOD-YOLO's post-processing."},{"cited_title":"You only look once: Unified, real-time object detection,","cited_arxiv_id":null,"evidence_quote":"Serves as a comparison baseline in Table 1 for detector performance on VisDrone2019-DET-val."}],"review_version":1}