{"id":"7bd97177-18e5-46ce-9f67-a08f1cd05dba","arxiv_id":"2504.18136","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"MASF-YOLO, an improved YOLOv11 with multi-scale aggregation, attention, and selective fusion modules, raises VisDrone2019 validation mAP@0.5 from 44.6% to 49.2%.","lead":"MASF-YOLO adds three feature-processing modules to the YOLOv11 detector and reports better accuracy on drone-view images from VisDrone2019. A generalist might read it to see how incremental architectural tweaks are being used to improve small-object detection from aerial platforms.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline gain over YOLOv11-m rests on single-run validation numbers; the test-set gap is only 0.7/0.2 mAP, so without error bars or tuned baselines the 'remarkable' claim is not yet established.","rationale":"The reader's weakest assumption is that the comparison is fair and not an artifact of tuning; I agree and sharpen this with a concrete, internally visible piece of evidence: the held-out test split in Table II shows the advantage over YOLOv11-m shrinking to 0.7 mAP@0.5 and only 0.2 mAP@0.5:0.95. This makes the 'remarkable' claim depend on a single-run validation gap that may be within noise or tuning sensitivity. The paper also provides no code, no seed variance, and omits the actual equations for MFAM (equations 1-6 are placeholders), so the empirical evidence cannot be checked independently. These issues do not prove the claim false, but they justify keeping the verdict conditional rather than accepting the results at face value. I therefore recommend no change to the reader's CONDITIONAL verdict.","tokens_in":8562,"tokens_out":10204,"duration_ms":105726,"concrete_test":"Train MASF-YOLO-s and YOLOv11-m on VisDrone2019 with at least 3 random seeds and a small hyperparameter grid (e.g., learning rate in {0.005, 0.01} and epochs in {100, 200}), then report mean and standard deviation on both the validation and test splits. If the best or mean YOLOv11-m test mAP@0.5:0.95 is within 0.3 of MASF-YOLO-s, or if the validation gap is not significant across seeds, the headline advantage is not robust.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV-B reports a single training recipe (SGD, lr 0.01, 100 epochs, batch 12) applied to all models, with no random seeds, no error bars, and no per-model hyperparameter search. The central comparison in Table II therefore equates one run of MASF-YOLO-s with one run of YOLOv11-m. On the validation set, MASF-YOLO-s leads by 1.4 mAP@0.5 (49.2 vs 47.8) and 0.7 mAP@0.5:0.95. But on the held-out test set, the same table shows the gap shrinks to 0.7 mAP@0.5 (42.8 vs 42.1) and 0.2 mAP@0.5:0.95 (26.8 vs 26.6). That 0.2-point gap is plausibly within run-to-run noise, especially since neither model is reported with multiple seeds. Also, because the same hyperparameters may favor the smaller baseline (larger models often need lower learning rates or longer schedules), the reported 1.4-point validation edge could be partly a tuning artifact. No code is released, and the MFAM equations (1)-(6) are not actually printed in the manuscript, so the experimental evidence cannot be independently reproduced.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes MASF-YOLO, a YOLOv11-based detector for small objects in drone-captured images. Four contributions are claimed: (i) a P2 high-resolution detection layer, (ii) a Multi-scale Feature Aggregation Module (MFAM) derived from PKINet, (iii) an Improved Efficient Multi-scale Attention (IEMA) module built on EMA, and (iv) a Dimension-Aware Selective Integration (DASI) module borrowed from HCF-Net, together with skip connections in the neck. Experiments on VisDrone2019 report ablation gains over a YOLOv11-s baseline and claim that MASF-YOLO-s outperforms YOLOv11-m on both validation and test sets while using about 60% of its parameters and 65% of its GFLOPs. A comparison with six prior detectors is also provided.","tokens_in":8835,"tokens_out":4113,"duration_ms":39376,"significance":"If the results are robust, the paper offers a practical recipe for improving small-object detection on drone views, with a clear accuracy/efficiency trade-off versus YOLOv11-m. The ablation is internally consistent, the parameter counts are reported coherently, and the methods are evaluated on the standard VisDrone2019 benchmark. However, the significance is limited by the fact that the main modules are adaptations of existing work (PKINet, EMA, HCF-Net), the MFAM equations are not actually printed, no code is released, and the experimental evidence rests on single training runs without error bars. The headline claim that MASF-YOLO-s surpasses YOLOv11-m may be fragile because the test-set gaps are very small (0.7 mAP@0.5 and 0.2 mAP@0.5:0.95).","major_comments":[{"comment":"The central claim that MASF-YOLO-s outperforms YOLOv11-m rests on a single training run per model. On the test set the gap is only 0.7 mAP@0.5 and 0.2 mAP@0.5:0.95, a difference that is plausibly within run-to-run noise for detectors on VisDrone2019. The paper reports no random seeds, no multiple runs, and no error bars or significance tests. I request that the authors repeat each experiment with at least three seeds and report mean and standard deviation, or otherwise temper the claim that MASF-YOLO-s is superior to YOLOv11-m.","section":"Section IV-B, Table II"},{"comment":"The MFAM module is the first named contribution, but the mathematical expressions referenced as Eqs. (1)-(6) are not present in the manuscript. Without these equations (or a precise textual description of the operations), the module cannot be reproduced or verified. No code is provided either. The authors should include the full equations or a pseudocode block for MFAM.","section":"Section III-A, Eqs. (1)-(6)"},{"comment":"The comparison with state-of-the-art detectors is sparse and mostly dated. Table III includes Faster R-CNN, Cascade R-CNN, EfficientDet, TPH-YOLOv5, YOLOv8-m, and MSFE-YOLO, but none of the recent drone-view detectors cited in the related work, such as FFCA-YOLO, LSKNet, PKINet, or HCF-Net, are compared. Given the abstract's claim of \"state-of-the-art\" performance, the authors should expand the comparison to include these more recent methods, or weaken the claim to \"competitive with selected baselines.\"","section":"Section IV-E, Table III"},{"comment":"The efficiency narrative is incomplete. The abstract emphasizes that MASF-YOLO-s requires only about 60% of the parameters and 65% of the GFLOPs of YOLOv11-m, which is true, but it does not disclose that relative to the direct baseline YOLOv11-s, MASF-YOLO-s more than doubles the GFLOPs (44.3 vs. 21.3) and adds 28% parameters. Moreover, the paper reports no inference speed (FPS or latency) on the actual hardware, so the practical \"light-weight\" and \"real-time\" claims for UAV deployment are not substantiated. Please report latency and FPS, and explicitly discuss the GFLOPs increase over the baseline.","section":"Abstract and Table II"},{"comment":"All models are trained with the same recipe (SGD, lr 0.01, momentum 0.937, 100 epochs, batch size 12) with no hyperparameter search per model. Larger models such as YOLOv11-m often require different learning rates or longer schedules, so the comparison in Table II may be biased in favor of the smaller MASF-YOLO-s. The authors should either perform a modest hyperparameter search per baseline or add a discussion of this limitation and justify the chosen recipe as standard for all models.","section":"Section IV-B"}],"minor_comments":[{"comment":"There are several typos and grammatical errors, e.g., \"superoiror over tranditional\" in the Introduction and \"UVA missions\" in Section I. The manuscript would benefit from a careful proofreading pass.","section":"Abstract and Introduction"},{"comment":"The text says \"Compared to the PKI Module [12], MKAM significantly enhances...\" but the module acronym is MFAM, not MKAM. Please correct this.","section":"Section III-A"},{"comment":"The subsection heading \"Training Sets\" is misleading; the content describes the experimental environment and training hyperparameters. Rename it to \"Experimental Setup\" or \"Training Details.\"","section":"Section IV-B"},{"comment":"The table lists MSFE-YOLO-l without a parameter count, and TPH-YOLOv5-s also lacks one. If these numbers are unavailable, add a note explaining the omission rather than leaving an empty cell.","section":"Table III"},{"comment":"The caption says \"MASF-YOLO-s are highlighted with red bounding boxes\" but the text is grammatically incomplete; also, the figure resolution in the PDF is low, making the small-target detection boxes hard to verify. Please provide higher-quality figures.","section":"Figure 5"}],"recommendation":"major_revision","confidential_remarks":"The paper is an incremental engineering contribution whose core modules are mostly borrowed from PKINet, EMA, and HCF-Net. The main risk is that the headline claim of superiority over YOLOv11-m is based on single-run numbers with a test-set gap of 0.2 mAP@0.5:0.95, which may be within noise. The missing MFAM equations are a serious reproducibility problem. The SOTA comparison is weak for a journal submission. With the requested revisions (error bars, full equations, expanded comparison, and fair efficiency discussion), the paper could become acceptable, but in its current form it does not meet the bar for a strong journal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: this is a standard “YOLO plus three borrowed modules” paper, and the authors are upfront that MFAM, IEMA, and DASI come from PKINet, EMA, and HCF-Net. The specific combination is new; the headline claim that MASF-YOLO-s beats YOLOv11-m with fewer parameters and GFLOPs is consistent with their tables. The ablation shows each module adds something on the VisDrone2019 validation set, and that monotonic trend is more than many papers in this genre give you.\n\nWhat it does well: it credits its sources, uses an external benchmark, reports both validation and test numbers, and gives parameter and FLOP counts. That is a solid baseline of practice.\n\nSoft spots. First, equations (1)–(6) are referenced in Section III-A but never actually appear in the manuscript. That alone stops anyone from reimplementing MFAM exactly. Second, every model is trained once, with one recipe, no seeds, no error bars. The test-set gap between MASF-YOLO-s and YOLOv11-m is 0.7 mAP@0.5 and 0.2 mAP@0.5:0.95. The latter is within run-to-run noise, so the word “remarkably” is doing more work than the data supports. Third, the efficiency story is selective: compared to YOLOv11-s, MASF-YOLO-s doubles GFLOPs (21.3 to 44.3). The comparison to YOLOv11-m is fair on params and compute, but practitioners choosing within the s-scale will feel the cost jump. Fourth, the SOTA table is sparse and a bit dated—no RT-DETR, no YOLOv9/10, and only one other drone-view small-object detector. Citations look fine, and there is no circularity.\n\nWho is this for? Practitioners working on drone-view detection who want a concrete YOLOv11 improvement recipe. It does not open a new direction, but it is a usable engineering data point.\n\nRecommendation: it deserves a serious referee rather than a desk reject, but only with major revision—add the missing equations, ideally multi-seed runs, and a broader SOTA comparison. As it stands, the central accuracy claim is conditional.","headline":"A conventional but honest YOLO-plus-modules paper whose efficiency claim over YOLOv11-m is plausible on params/FLOPs, though the accuracy edge is not yet established because of single runs and missing equations.","tokens_in":9315,"tokens_out":2323,"would_cite":false,"duration_ms":23190,"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":"Small drone detector beats a larger YOLO at 60 percent of the cost.","keywords":["small object detection","drone view","UAV imagery","YOLOv11","multi-scale feature fusion","attention module","VisDrone2019","lightweight detector"],"falsifier":"Re-run Table II with a hyperparameter sweep or multiple random restarts for each model on the same VisDrone2019 splits. If a tuned or repeated-seed YOLOv11-s or YOLOv11-m reaches or exceeds MASF-YOLO-s's 49.2% mAP@0.5, the central accuracy advantage collapses; likewise, removing MFAM, IEMA, or DASI should measurably reduce performance if the ablation story is right.","tokens_in":8385,"feed_emoji":"🚁","tokens_out":4950,"duration_ms":41549,"temperature":0.7,"pith_summary":"This paper claims that a small object detector for drone imagery can be made substantially more accurate by adding four targeted components to YOLOv11: a fine-grained P2 detection layer, a multi-scale feature aggregation module, a lightweight attention module, and an adaptive fusion module. On the VisDrone2019 benchmark, the resulting MASF-YOLO-s raises mAP@0.5 from 44.6% to 49.2% over YOLOv11-s and edges past YOLOv11-m (47.8%) while using about 60% of its parameters and 65% of its computation. The authors argue that the gain matters because drone-view small objects are tiny, scale-variable, and embedded in background noise, and they present an ablation table showing that each added module contributes to the improvement.","feed_headline":"Drone detector beats a larger YOLO at 60 percent of the cost","feed_subtitle":"On VisDrone2019, four small-object upgrades push YOLOv11-small past YOLOv11-medium in accuracy.","key_machinery":"The central mechanism is a set of four plug-in modules acting on YOLOv11's backbone and neck. MFAM replaces a large-kernel convolution with two strip convolutions (1×k and k×1) run in parallel at k=7 and k=9, plus a parallel 3×3 depthwise path, all summed with the input to gather multi-scale context cheaply. IEMA splits channels into groups, runs parallel sub-networks with 3×3 and 1×5/5×1 depthwise convolutions plus an identity path, and uses cross-spatial learning to suppress background noise. DASI aligns low- and high-dimensional features to the current layer, splits channels into partitions, and uses sigmoid-activated weights from the current layer to fuse them, with a residual connection. A P2 layer preserves high-resolution spatial detail, and skip connections in the neck retain shallow semantic information. Together these modules are the claimed reason a small model can match or beat a larger one.","core_discovery":"On its own terms, the central result is an accuracy-efficiency trade-off: MASF-YOLO-s reaches 49.2% mAP@0.5 and 32.9% mAP@0.5:0.95 on the VisDrone2019 validation set, surpassing YOLOv11-m (47.8% and 32.2%) with 12.05M parameters and 44.3 GFLOPs versus 20.04M and 67.7. The same pattern appears on the test set, where MASF-YOLO-s scores 42.8% mAP@0.5 versus 42.1% for YOLOv11-m. The paper traces the improvement to the P2 fine-grained feature layer, a Multi-scale Feature Aggregation Module (MFAM) built from parallel multi-scale depthwise convolutions, an Improved Efficient Multi-scale Attention (IEMA) module, and a Dimension-Aware Selective Integration (DASI) module, and supports this with an ablation study showing rising accuracy as each module is added to the baseline.","pith_inferences":["A natural extension is to test the same four-module recipe on other small-object benchmarks, such as wider aerial or remote sensing datasets, to see whether the P2 layer and fusion modules generalize beyond VisDrone2019.","Because the paper reports a single training recipe with no repeated runs, comparing tuned baselines under an equal compute budget would separate architectural gains from tuning advantages; that is a testable next step the paper does not perform.","The modular design suggests that other detectors could adopt MFAM, IEMA, and DASI independently, since each module targets a distinct bottleneck: resolution, context, noise suppression, and scale fusion."],"forward_implications":["If the claims hold, a drone-view detector can gain roughly 4.6 mAP@0.5 points over YOLOv11-s without scaling up to a medium-sized model.","Deployment scenarios with tight memory or latency budgets could use MASF-YOLO-s instead of YOLOv11-m, keeping most of the accuracy gain at about 60% of the parameters.","The ablation table implies the four modules address complementary failure modes, so removing any single module should measurably reduce performance.","The same recipe applied at the n-scale lifts MASF-YOLO-n above YOLOv11-s, suggesting the improvements transfer across model capacities."],"supporting_citations":[{"why":"Supplies the YOLOv11 baseline network that MASF-YOLO modifies.","marker":"[9]"},{"why":"Provides the PKINet module that MFAM optimizes into a lighter multi-scale aggregation block.","marker":"[12]"},{"why":"Provides the EMA attention module that IEMA is inspired by and extends with multi-scale depthwise convolutions.","marker":"[16]"},{"why":"Supplies the DASI fusion mechanism that the paper adaptively fuses low- and high-dimensional features.","marker":"[17]"},{"why":"Serves as a state-of-the-art single-stage detector baseline in the comparison table.","marker":"[8]"},{"why":"Provides a drone-captured YOLOv5 baseline that the paper compares against for small-object performance.","marker":"[24]"},{"why":"Gives a recent drone-view YOLOv8-based detector whose reported accuracy serves as a competitive baseline.","marker":"[25]"},{"why":"Offers an efficiency-oriented detector baseline used in the state-of-the-art comparison.","marker":"[20]"}],"fun_headline_variants":["Tiny YOLO tops a bigger model on drone images","Small drone-detector beats a bigger YOLO at 60% cost","Four upgrades let tiny YOLO outdo big YOLO on drones","Efficient drone detection: small model, big win","Less compute, more accuracy for drone object detection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the comparison is fair: every model is trained with the same fixed recipe (SGD, learning rate 0.01, 100 epochs, batch size 12, 640×640 input) with no hyperparameter search or repeated runs, so if the baseline YOLOv11 models are under-tuned, the reported advantage could come from tuning rather than from the new modules.","fun_headline_variants_meta":{"raw":{"variants":["Tiny YOLO tops a bigger model on drone images","Small drone-detector beats a bigger YOLO at 60% cost","Four upgrades let tiny YOLO outdo big YOLO on drones","Efficient drone detection: small model, big win","Less compute, more accuracy for drone object detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001117,"raw_usage":{"total_tokens":4746,"prompt_tokens":1135,"completion_tokens":3611,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":751,"completion_tokens_details":{"reasoning_tokens":3524}},"tokens_in":751,"tokens_out":3611,"duration_ms":24857,"temperature":1.0,"reasoning_tokens":3524,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:22:40.277501+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run Table II with a hyperparameter sweep or multiple random restarts for each model on the same VisDrone2019 splits. If a tuned or repeated-seed YOLOv11-s or YOLOv11-m reaches or exceeds MASF-YOLO-s's 49.2% mAP@0.5, the central accuracy advantage collapses; likewise, removing MFAM, IEMA, or DASI should measurably reduce performance if the ablation story is right.","supporting_citations":[{"cited_title":"Poly Kernel Inception Network for Remote Sensing Detection","cited_arxiv_id":"2403.06258","evidence_quote":"Provides the PKINet module that MFAM optimizes into a lighter multi-scale aggregation block."},{"cited_title":"Efficient Multi-Scale Attention Module with Cross-Spatial Learning","cited_arxiv_id":"2305.13563","evidence_quote":"Provides the EMA attention module that IEMA is inspired by and extends with multi-scale depthwise convolutions."},{"cited_title":"Jocher , A","cited_arxiv_id":null,"evidence_quote":"Serves as a state-of-the-art single-stage detector baseline in the comparison table."},{"cited_title":"TPH-YOLOv5: Improved YOLOv5 based on transformer prediction head for object detection on drone-captured scenarios","cited_arxiv_id":null,"evidence_quote":"Provides a drone-captured YOLOv5 baseline that the paper compares against for small-object performance."},{"cited_title":"MSFE-YOLO: An Improved YOLOv8 Network for Object Detection on Drone View","cited_arxiv_id":null,"evidence_quote":"Gives a recent drone-view YOLOv8-based detector whose reported accuracy serves as a competitive baseline."},{"cited_title":"EfficientDet: Scalable and Efficient Object Detection","cited_arxiv_id":null,"evidence_quote":"Offers an efficiency-oriented detector baseline used in the state-of-the-art comparison."}],"review_version":1}