REVIEW 4 major objections 5 minor 25 references
SOD-YOLO: Enhancing YOLO-Based Detection of Small Objects in UAV Imagery
T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash
Pith's one-line read 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.
desk verdict 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. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
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.
What would settle it
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.
Extended reading notes
Core claim
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.
Load-bearing premise
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.
Editorial extensions
If this is right
- 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.
Reading between the lines
- 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.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
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.
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 (4)
- [Table 2, §4.6] 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.
- [Tables 1 and 2] 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.
- [§4.3] 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.
- [§4.3, Table 1] 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.
minor comments (5)
- [§4.5] 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.
- [References] 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.
- [Figure 1] 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.
- [§3.1] 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.
- [Throughout] 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).
Circularity Check
No circularity: the reported gains are measured empirical outcomes from ablations, not derived from fitted inputs or self-cited constraints.
full rationale
SOD-YOLO is an empirical engineering paper. Its central claim is that the proposed detector achieves higher mAP on VisDrone2019-DET than a YOLOv8-m baseline. That claim is supported by direct measurements in Table 1 and the ablation study in Table 2, where each component (ASF, P2, Soft-NMS) is added one at a time and evaluated on the same validation set. There is no derivation chain in which an equation for the mAP gain is constructed from fitted parameters; the gains are measured outcomes. Soft-NMS is a standard post-processing method cited to [18], and its contribution in Table 2 is a measured change in evaluation output using the same model weights, not a predicted quantity obtained from a fitted parameter. The ASF mechanism is explicitly attributed to prior work [15]; the paper's use of the term 'proposed ScalSeq' is lax but the citation makes the external origin clear. No self-citation is load-bearing, since the authors cite no prior work of their own as justification. The minor inconsistency between 0.351 (Table 1) and 0.352 (Table 2) and the absence of NMS threshold details are reproducibility and correctness concerns, not evidence of circularity. Under the given rules, a non-finding with score 0 is appropriate because no claim reduces by construction to its inputs.
Assumptions & free parameters
assumptions (3)
- domain assumption VisDrone2019-DET annotations are treated as reliable ground truth and mAP is the appropriate evaluation metric.
- domain assumption The ASF mechanism from ASF-YOLO [15] works as described and transfers to YOLOv8-m without degradation.
- domain assumption All models are trained under identical settings and the single-run baseline numbers are representative.
Cite this review
Pith. "Pith review of SOD-YOLO: Enhancing YOLO-Based Detection of Small Objects in UAV Imagery." pith.science (2026). https://pith.science/paper/HOAKXTTI
@misc{pith2026250712727,
author = {Pith},
title = {Pith review of: SOD-YOLO: Enhancing YOLO-Based Detection of Small Objects in UAV Imagery},
year = {2026},
howpublished = {\url{https://pith.science/paper/HOAKXTTI}},
note = {Machine review of arXiv:2507.12727}
}
abstract
Small object detection remains a challenging problem in the field of object detection. To address this challenge, we propose an enhanced YOLOv8-based model, SOD-YOLO. This model integrates an ASF mechanism in the neck to enhance multi-scale feature fusion, adds a Small Object Detection Layer (named P2) to provide higher-resolution feature maps for better small object detection, and employs Soft-NMS to refine confidence scores and retain true positives. Experimental results demonstrate that SOD-YOLO significantly improves detection performance, achieving a 36.1% increase in mAP$_{50:95}$ and 20.6% increase in mAP$_{50}$ on the VisDrone2019-DET dataset compared to the baseline model. These enhancements make SOD-YOLO a practical and efficient solution for small object detection in UAV imagery. Our source code, hyper-parameters, and model weights are available at https://github.com/iamwangxiaobai/SOD-YOLO.
Figures
Reference graph
Works this paper leans on
-
[10]
You only look once: Unified, real-time object detection,
J. Redmon, et al., “You only look once: Unified, real-time object detection,” in Pro- ceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016, pp. 779-788
work page 2016
-
[19]
Session Peering Provisioning Framework (SPPF),
K. Cartwright, et al. , “Session Peering Provisioning Framework (SPPF),” 2016
work page 2016
-
[1]
From unmanned systems to autonomous intelligent systems,
J. Chen, J. Sun, and G. Wang, “From unmanned systems to autonomous intelligent systems,” Engineering, vol. 12, 2022, pp. 16-19
work page 2022
-
[2]
Rich feature hierarchies for accu- rate object detection and semantic segmentation,
R. Girshick, J. Donahue, T. Darrell, and J. Malik, “Rich feature hierarchies for accu- rate object detection and semantic segmentation, ” in CVPR, 2014, pp. 580-587
work page 2014
-
[3]
Microsoft COCO: common objects in context,
T. Lin, et al. , “Microsoft COCO: common objects in context, ” in ECCV, Cham: Springer, 2014, pp. 740-755
work page 2014
-
[4]
SSD: single shot multibox detector,
W. Liu, et al. , “SSD: single shot multibox detector, ” in ECCV, 2016, pp. 21-37
work page 2016
-
[5]
Faster R-CNN: towards real-time object detection with region proposal networks,
S. Ren, et al. , “Faster R-CNN: towards real-time object detection with region proposal networks, ” IEEE Transactions on Pattern Analysis & Machine Intelligence, vol. 39, no. 6, 2017, pp. 1137-1149
work page 2017
-
[6]
You only look once: unified, real-time object detection,
J. Redmon, et al. , “You only look once: unified, real-time object detection, ” in CVPR, 2016, pp. 779-788
work page 2016
Show all 25 references
-
[7]
YOLO9000: better, faster, stronger,
J. Redmon, and A. Farhadi, “YOLO9000: better, faster, stronger, ” in CVPR, 2017, pp. 7263-7271
2017
-
[8]
YOLOv3: an incremental improvement,
J. Redmon, and A. Farhadi, “YOLOv3: an incremental improvement,” arXiv: 1804.02767, 2018
2018 arXiv
-
[9]
YOLOv4: Optimal speed and accuracy of object detection,
A. Bochkovskiy, C. Y. Wang, and H. Y. M. Liao, “YOLOv4: Optimal speed and accuracy of object detection,” arXiv preprint arXiv: 2004.10934, 2020
2004 arXiv
-
[11]
YOLOv7: Trainable bag-of- freebies sets new state-of-the-art for real-time object detectors,
C. Y. Wang, A. Bochkovskiy, and H. Y. M. Liao, “YOLOv7: Trainable bag-of- freebies sets new state-of-the-art for real-time object detectors, ” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 7464-7475. 12
2023
-
[12]
YOLOv9: Learning What You Want to Learn Using Pro- grammable Gradient Information,
C. Y. Wang, et al. , “YOLOv9: Learning What You Want to Learn Using Pro- grammable Gradient Information,” arXiv preprint arXiv: 2402.13616, 2024
2024 arXiv
-
[13]
YOLOv10: Real-time end-to-end object detection,
A. Wang, et al. , “YOLOv10: Real-time end-to-end object detection,” arXiv preprint arXiv: 2405.14458, 2024
2024 arXiv
-
[14]
VisDrone-DET2019: The vision meets drone object detection in im- age challenge results,
D. Du, et al., “VisDrone-DET2019: The vision meets drone object detection in im- age challenge results,” in Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, 2019
2019
-
[15]
ASF-YOLO: A novel YOLO model with attentional scale sequence fusion for cell instance segmentation,
M. Kang, et al. , “ASF-YOLO: A novel YOLO model with attentional scale sequence fusion for cell instance segmentation,” Image and Vision Computing, vol. 147, 2024, 105057
2024
-
[16]
CSPNet: A new backbone that can enhance learning capability of CNN,
C. Y. Wang, et al. , “CSPNet: A new backbone that can enhance learning capability of CNN,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2020, pp. 390-391
2020
-
[17]
EfficientNet: Rethinking model scaling for convolutional neural networks,
M. Tan, and Q. Le, “EfficientNet: Rethinking model scaling for convolutional neural networks, ” in International Conference on Machine Learning, PMLR, 2019, pp. 6105- 6114
2019
-
[18]
Soft-NMS: improving object detection with one line of code,
N. Bodla, et al. , “Soft-NMS: improving object detection with one line of code,” in Proceedings of the IEEE International Conference on Computer Vision, 2017, pp. 5561-5569
2017
-
[20]
EdgeYOLO: An edge-real-time object detector,
S. Liu, et al., “EdgeYOLO: An edge-real-time object detector,” in 2023 42nd Chinese Control Conference (CCC), IEEE, 2023, pp. 7507-7512
2023
-
[21]
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, Springer International Publishing, 2016, pp. 21–37
2016
-
[22]
Object detection with deep learning: A review,
Z.-Q. Zhao, P. Zheng, S.-T. Xu, and X. Wu, “Object detection with deep learning: A review,” arXiv preprint arXiv:1807.05511, 2019. [Online]. Available: https://arxiv.org/abs/1807.05511
2019 arXiv
-
[23]
Perceptual Generative Adversar- ial Networks for Small Object Detection,
J. Li, X. Liang, Y. Wei, T. Xu, J. Feng, and S. Yan, “Perceptual Generative Adversar- ial Networks for Small Object Detection,” IEEE Transactions on Image Processing, 2017
2017
-
[24]
The Unmanned Aerial Vehicle Bench- mark: Object Detection and Tracking,
D. Du, Y. Qi, H. Yu, Y. Yang, and K. Duan, “The Unmanned Aerial Vehicle Bench- mark: Object Detection and Tracking,” in Springer, Cham, 2018
2018
-
[25]
Efficient Non-Maximum Suppression,
A. Neubeck and L. J. Van Gool, “Efficient Non-Maximum Suppression,” inProceedings of the International Conference on Pattern Recognition (ICPR), 2006. 13
2006
Reviewed August 6, 2026 · model on record in the stance chip above.
Discussion (0). Sign in to comment.