Pith. sign in

REVIEW 5 major objections 5 minor 1 cited by

MASF-YOLO: An Improved YOLOv11 Network for Small Object Detection on Drone View

T0 review · 5 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Small drone detector beats a larger YOLO at 60 percent of the cost.

desk verdict 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. read the letter →

arxiv 2504.18136 v1 pith:STVAHVSS submitted 2025-04-25 cs.CV

Liugang Lu , Dabin He , Congxiang Liu , Zhixiang Deng This is my paper · ORCID
classification cs.CV
keywords smallobjectdetectiondroneviewUAVimageryYOLOv11multi-scalefeaturefusionattentionmoduleVisDrone2019lightweightdetector
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Editorial extensions

If this is right

  • 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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • 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.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 5 minor

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.

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 (5)
  1. [Section IV-B, Table II] 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.
  2. [Section III-A, Eqs. (1)-(6)] 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.
  3. [Section IV-E, Table III] 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."
  4. [Abstract and Table II] 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.
  5. [Section IV-B] 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.
minor comments (5)
  1. [Abstract and Introduction] 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.
  2. [Section III-A] The text says "Compared to the PKI Module [12], MKAM significantly enhances..." but the module acronym is MFAM, not MKAM. Please correct this.
  3. [Section IV-B] The subsection heading "Training Sets" is misleading; the content describes the experimental environment and training hyperparameters. Rename it to "Experimental Setup" or "Training Details."
  4. [Table III] 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.
  5. [Figure 5] 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.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: empirical evaluation on the external VisDrone2019 benchmark with modules cited from independent prior work; the claimed metrics are measured, not fitted or derived from the method's own assumptions.

full rationale

This paper makes an empirical architecture-improvement claim: augmenting YOLOv11 with P2 layers, MFAM, IEMA, and DASI improves mAP on VisDrone2019. The reported numbers are obtained by training on an external dataset and evaluating on its validation and test splits; there is no mathematical derivation in which an output is defined in terms of the input, and no fitted parameter is relabeled as a prediction. The ablation table (Table I) compares measured mAP values after successively adding modules to a baseline, so each row is an independent experimental outcome rather than a logical consequence of a definition. The comparisons in Tables II and III are against YOLOv11 variants and other detectors reported in the same experimental setup. The paper does cite prior work for inspiration (PKINet, EMA, HCF-Net), but those citations are external and load-bearing only as design antecedents; they do not establish the claimed quantitative results. Concerns about single-seed training, hyperparameter tuning, or missing module equations are validity and reproducibility issues, not circularity. No self-citation chain, no uniqueness theorem, and no ansatz smuggled via citation appears in the provided text. Therefore the score is 0.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

No free parameters or invented entities. The paper depends on standard deep learning training and evaluation assumptions, and on the representativeness of the VisDrone2019 benchmark.

assumptions (3)
  • domain assumption VisDrone2019 is a representative benchmark for drone-view small object detection
    The paper uses VisDrone2019 throughout without justifying its generality to other drone datasets.
  • standard math mAP@0.5 and mAP@0.5:0.95 are appropriate metrics for comparing detectors
    These are standard metrics in object detection, accepted by the community.
  • domain assumption The training recipe (SGD, initial LR 0.01, 100 epochs, batch 12) is equally suitable for all compared models
    The paper uses the same recipe for baselines and variants without showing per-model tuning, which is a common but unverified assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MASF-YOLO: An Improved YOLOv11 Network for Small Object Detection on Drone View." pith.science (2026). https://pith.science/paper/STVAHVSS

@misc{pith2026250418136,
  author       = {Pith},
  title        = {Pith review of: MASF-YOLO: An Improved YOLOv11 Network for Small Object Detection on Drone View},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/STVAHVSS}},
  note         = {Machine review of arXiv:2504.18136}
}
read the original abstract

With the rapid advancement of Unmanned Aerial Vehicle (UAV) and computer vision technologies, object detection from UAV perspectives has emerged as a prominent research area. However, challenges for detection brought by the extremely small proportion of target pixels, significant scale variations of objects, and complex background information in UAV images have greatly limited the practical applications of UAV. To address these challenges, we propose a novel object detection network Multi-scale Context Aggregation and Scale-adaptive Fusion YOLO (MASF-YOLO), which is developed based on YOLOv11. Firstly, to tackle the difficulty of detecting small objects in UAV images, we design a Multi-scale Feature Aggregation Module (MFAM), which significantly improves the detection accuracy of small objects through parallel multi-scale convolutions and feature fusion. Secondly, to mitigate the interference of background noise, we propose an Improved Efficient Multi-scale Attention Module (IEMA), which enhances the focus on target regions through feature grouping, parallel sub-networks, and cross-spatial learning. Thirdly, we introduce a Dimension-Aware Selective Integration Module (DASI), which further enhances multi-scale feature fusion capabilities by adaptively weighting and fusing low-dimensional features and high-dimensional features. Finally, we conducted extensive performance evaluations of our proposed method on the VisDrone2019 dataset. Compared to YOLOv11-s, MASFYOLO-s achieves improvements of 4.6% in mAP@0.5 and 3.5% in mAP@0.5:0.95 on the VisDrone2019 validation set. Remarkably, MASF-YOLO-s outperforms YOLOv11-m while requiring only approximately 60% of its parameters and 65% of its computational cost. Furthermore, comparative experiments with state-of-the-art detectors confirm that MASF-YOLO-s maintains a clear competitive advantage in both detection accuracy and model efficiency.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. SDG-Track: A Heterogeneous Observer-Follower Framework for High-Resolution UAV Tracking on Embedded Platforms

    cs.CV 2025-12 conditional novelty 5.0 of 10

    A heterogeneous observer-follower tracker achieves ~35 FPS on a Jetson Orin Nano at 1080p with ~3% precision loss versus frame-by-frame YOLO11l detection on UAV tracking benchmarks.

Reference graph

Works this paper leans on

25 extracted references · 23 canonical work pages · cited by 1 Pith paper

  1. [1]

    Machine Learning for Precision Agriculture Using Imagery from Unmanned Aerial Vehicles (UAVs): A Survey

    I. Zualkernan, D. A. Abuhani, M. H. Hussain, J Khan, and M. ElMohandes. "Machine Learning for Precision Agriculture Using Imagery from Unmanned Aerial Vehicles (UAVs): A Survey. " Drones ,vol. 7, p. 382, 2023

    work page 2023

  2. [2]

    A review on early wildfire detection from unmanned aerial vehicles using deep learning-based computer vision algorithms

    A. Bouguettaya, H. Zarzour, A.M Taberkit, and A. Kechida. "A review on early wildfire detection from unmanned aerial vehicles using deep learning-based computer vision algorithms ." Signal Processing, vol. 190, 2022

    work page 2022

  3. [3]

    Smart Traffic Monitoring through Drone Images via Yolov5 and Kalman Filter

    M. Hanzla, S. Ali and A. Jalal, "Smart Traffic Monitoring through Drone Images via Yolov5 and Kalman Filter." In 2024 5th International Conference on Advancements in Computational Sciences (ICACS), Lahore, Pakistan, 2024, pp. 1–8

    work page 2024

  4. [4]

    Fast R-CNN

    R. Girshick. "Fast R-CNN." In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2015, pp. 1440–1448

    work page 2015

  5. [5]

    Faster R-CNN: Towards real- time object detection with region proposal networks

    S. Ren, K. He, R. Girshick, and J. Sun. "Faster R-CNN: Towards real- time object detection with region proposal networks ." IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 39, no. 6, pp. 1137–1149, 2017

    work page 2017

  6. [6]

    Mask R-CNN

    K. He, G. Gkioxari, P. Dollár, and R. Girshick. “Mask R-CNN.” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 42, no. 2, pp. 386–397, 2020

    work page 2020

  7. [7]

    Jocher (2022) Yolov5

    G. Jocher (2022) Yolov5. [Online]. Available: https:// github.com/ ultralytics/yolov5

    work page 2022

  8. [8]

    Jocher , A

    G. Jocher , A. Chaurasia, and J. Qiu. (2023). Yolov8. [Online]. Available: https://github.com/ ultralytics/ultralytics

    work page 2023

Show all 25 references
  1. [9]

    YOLOv11: An Overview of the Key Architectural Enhancements

    R. Khanam and M. Hussain. "YOLOv11: An Overview of the Key Architectural Enhancements." arXiv preprint arXiv:2410.17725, 2024

    2024 arXiv

  2. [10]

    SSD: Single shot multibox detector,

    W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. Reed, C. -Y. Fu, and A. C. Berg, "SSD: Single shot multibox detector," in Computer Vision – ECCV 2016, Amsterdam, The Netherlands, 2016, pp. 21–37

    2016

  3. [11]

    Large Selective Kernel Network for Remote Sensing Object Detection

    Y. Li, Q . Hou, Z . Zheng, M . Cheng, J . Yang, and X. Li. "Large Selective Kernel Network for Remote Sensing Object Detection." arXiv preprint arXiv:2303.09030, 2023

    2023 arXiv

  4. [12]

    Poly Kernel Inception Network for Remote Sensing Detection

    X. Cai, Q. Lai, Y. Wang, W. Wang, Z. Sun, and Y. Yao. "Poly Kernel Inception Network for Remote Sensing Detection. " arXiv preprint arXiv:2403.06258,2024

    2024 arXiv

  5. [13]

    High-Resolution Feature Pyramid Network for Small Object Detection on Drone View

    Z. Chen, H. Ji, Y. Zhang, Z. Zhu, and Y. Li. "High-Resolution Feature Pyramid Network for Small Object Detection on Drone View ." IEEE Transactions on Circuits and Systems for Video Technology , vol. 34, no. 1, pp. 475–489, 2024

    2024

  6. [14]

    FFCA-YOLO for small object detection in remote sensing images

    Y. Zhang, M. Ye, G. Zhu, Y. Liu, P. Guo, and J. Yan. "FFCA-YOLO for small object detection in remote sensing images." IEEE Transactions on Geoscience and Remote Sensing ,vol. 62, pp. 1– 15, 2024

    2024

  7. [15]

    Visual attention network

    M. Guo, C . Lu, Z. Liu, M. Cheng, and S. Hu . "Visual attention network." Computational Visual Media, vol. 9, pp. 733–752, 2023

    2023

  8. [16]

    Efficient Multi-Scale Attention Module with Cross-Spatial Learning

    D. Ouyang, S. He, G. Zhang, M. Luo, H. Guo, J. Zhan, and Z. Huang. "Efficient Multi-Scale Attention Module with Cross-Spatial Learning." arXiv preprint arXiv:2305.13563,2023

    2023 arXiv

  9. [17]

    HCF-Net: Hierarchical Context Fusion Network for Infrared Small Object Detection

    S. Xu, S. Zheng, W. Xu, R. Xu, C. Wang, J. Zhang, X. Teng, A. Li, and L. Guo. "HCF-Net: Hierarchical Context Fusion Network for Infrared Small Object Detection." arXiv preprint arXiv:2403.10778,2024

    2024 arXiv

  10. [18]

    Feature Pyramid Networks for Object Detection

    T. -Y. Lin, P. Dollár, R. Girshick, K. He, B. Hariharan, and S. Belongie. "Feature Pyramid Networks for Object Detection." in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017, pp. 936–944

    2017

  11. [19]

    Path Aggregation Network for Instance Segmentation

    S. Liu, L. Qi, H. Qin, J. Shi, and J. Jia. "Path Aggregation Network for Instance Segmentation." in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018, pp. 8759– 8768

    2018

  12. [20]

    EfficientDet: Scalable and Efficient Object Detection

    M. Tan, R. Pang, and Q. V. Le. "EfficientDet: Scalable and Efficient Object Detection." in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020, pp. 10778– 10787

    2020

  13. [21]

    AFPN: Asymptotic Feature Pyramid Network for Object Detection

    G. Yang, J . Lei, Z . Zhu, S . Cheng, Z . Feng, and R. Liang. “AFPN: Asymptotic Feature Pyramid Network for Object Detection.” arXiv preprint arXiv:2306.15988,2023

    2023 arXiv

  14. [22]

    InceptionNeXt: When Inception Meets ConvNeXt

    W. Yu, P . Zhou, S . Yan, and X. Wang. "InceptionNeXt: When Inception Meets ConvNeXt." in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 5672–5683

    2024

  15. [23]

    Cascade R -CNN: Delving into High Quality Object Detection

    Z. Cai and N. Vasconcelos, "Cascade R -CNN: Delving into High Quality Object Detection." in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2018, pp. 6154–6162

    2018

  16. [24]

    TPH-YOLOv5: Improved YOLOv5 based on transformer prediction head for object detection on drone-captured scenarios

    X. Zhu, S. Lyu, X. Wang, and Q. Zhao. "TPH-YOLOv5: Improved YOLOv5 based on transformer prediction head for object detection on drone-captured scenarios ." in Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), 2021, pp. 2778–2788

    2021

  17. [25]

    MSFE-YOLO: An Improved YOLOv8 Network for Object Detection on Drone View

    S. Qi, X. Song, T. Shang, X. Hu , and K. Han . "MSFE-YOLO: An Improved YOLOv8 Network for Object Detection on Drone View ." IEEE Geoscience and Remote Sensing Letters, vol. 21, pp. 1-5, 2024

    2024

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.