Pith. sign in

REVIEW 3 major objections 5 minor 22 references

YOLO-ROC: A High-Precision and Ultra-Lightweight Model for Real-Time Road Damage Detection

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read YOLO-ROC, a 0.89M-parameter road damage detector, reaches 67.6% mAP50 and beats a YOLOv8n baseline three times its size.

desk verdict A useful efficiency benchmark for drone-based road damage detection, but the central attribution of the accuracy gain to the BMS-SPPF module is not supported by the ablation design. read the letter →

arxiv 2507.23225 v1 pith:IVQ7WK4K submitted 2025-07-31 cs.CV

Zicheng Lin , Weichao Pan This is my paper
classification cs.CV
keywords roaddamagedetectionYOLO-ROClightweightobjectBMS-SPPFmulti-scalefeatureextractionchannelcompressionsmall-targetRDD2022dataset
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 road damage detector built on YOLOv8 can be made about 70% smaller in parameters and 68% cheaper in computation while still improving detection accuracy, provided the compression is paired with a purpose-built attention module. The proposed model, YOLO-ROC, reaches 67.6% mean average precision at IoU 0.5 (mAP50) on the drone-captured RDD2022_China_Drone dataset, versus 66.2% for the 3.01M-parameter YOLOv8n baseline, while shrinking to 0.89M parameters, 2.6 GFLOPs, and a 2.0 MB model. The authors attribute the gain to their Bidirectional Multi-scale Spatial Pyramid Pooling Fast (BMS-SPPF) module, which is designed to capture fine cracks and small potholes that standard SPPF misses, and to a hierarchical channel compression strategy that halves backbone width. If the claim holds, real-time road inspection on drones and mobile devices can expect higher small-target accuracy from a fraction of the compute.

What carries the argument

The central object is the BMS-SPPF module (Bidirectional Multi-scale Spatial Pyramid Pooling Fast), a drop-in replacement for the SPPF layer in the YOLOv8 backbone. It chains the standard SPPF pooling branch into three attention stages: Multi-Scale Spatial Attention (MSSA) applies 1D depthwise convolutions with kernels of sizes 3, 5, 7, 9 along horizontal and vertical axes to capture elongated crack-like features; Channel Attention Preparation (CAP) condenses the feature map either by pooling or by lossless spatial-to-channel block recombination; and Multi-Head Channel Self-Attention (MHSA) computes scaled dot-product attention across channels to recalibrate feature importance and suppress background noise. The module's job is to let the heavily compressed network retain fine-grained spatial detail, particularly for small potholes (class D40), that naive channel narrowing would discard. The companion mechanism is the hierarchical compression itself: backbone max channels 1024 to 512 and C2f repeats (3,6,6,3) to (2,3,3,2), which drives the parameter and FLOP reduction.

What would settle it

Train the compressed YOLOv8n (max backbone channels 512, C2f repeats 2-3-3-2) with the standard SPPF module on RDD2022_China_Drone under the same 300-epoch SGD schedule and data augmentation. If that variant's mAP50 reaches or exceeds 67.6%, or its D40 mAP50 already shows a roughly 16.8% improvement, then the gains attributed to BMS-SPPF are actually produced by the compression; if it stays near the 66.2% baseline while YOLO-ROC holds 67.6%, the module's contribution is confirmed.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that an aggressive channel-compression scheme (backbone max width cut from 1024 to 512, C2f repetition counts reduced from (3,6,6,3) to (2,3,3,2), head channels halved) does not have to cost accuracy when the replaced SPPF module is upgraded to BMS-SPPF. The new module keeps the standard SPPF pooling chain, then passes its output through multi-scale spatial attention, a channel-attention preparation stage, and multi-head channel self-attention, which together suppress background interference and preserve the fine-grained morphology of cracks and potholes. The result, on the RDD2022_China_Drone dataset, is a 0.89M-parameter, 2.6-GFLOP, 2.0-MB model that reaches 67.6% mAP50 (2.11% above YOLOv8n), 39.7% mAP50:95, and a claimed 16.8% relative improvement on the small-target D40 category, while running at 142.74 FPS inference. On the RDD2022_China_Motorbike dataset the same model stays at 86.6% mAP50, within 0.3 points of the baseline with roughly a third of the parameters.

Load-bearing premise

The load-bearing premise is that the accuracy gain over YOLOv8n comes from the BMS-SPPF module itself, yet the paper's Table 5 changes channel width and C2f repetition counts in the same step as the SPPF replacement, so no experiment isolates the module's contribution.

Editorial extensions

If this is right

  • On RDD2022_China_Drone, YOLO-ROC at 0.89M parameters and 2.0 MB reaches 67.6% mAP50, a 2.11% relative gain over the 3.01M-parameter YOLOv8n baseline.
  • The small pothole class D40 is claimed to gain 16.8% in mAP50, making the smallest-target regime the main beneficiary of the attention design.
  • On RDD2022_China_Motorbike, the same compressed model scores 86.6% mAP50, within 0.3 points of YOLOv8n while using about a third of the parameters, indicating the compression transfers across capture platforms.
  • At 142.74 FPS inference throughput and 2.6 GFLOPs, the model stays in real-time territory while cutting computation by roughly two-thirds, the practical precondition for drone- or phone-mounted road surveys.

Reading between the lines

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

  • A controlled ablation that keeps the standard SPPF at the compressed width (max channels 512, C2f repeats 2-3-3-2) would settle whether the 2.11% gain is attributable to BMS-SPPF or to the compression itself, since no such run appears in Table 5.
  • Because the 16.8% D40 improvement is reported only in prose, publishing per-class AP on both datasets would let practitioners verify that the attention module helps the hardest class and not merely reweights the confusion matrix.
  • If BMS-SPPF is the true driver, the module should transfer to other YOLO-family backbones and to UAV inspection tasks beyond roads (power lines, facades), since its MSSA and MHSA stages make no road-specific assumptions.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes YOLO-ROC, a lightweight modification of YOLOv8n for road damage detection on the public RDD2022 dataset. The two main contributions are a Bidirectional Multi-scale Spatial Pyramid Pooling Fast (BMS-SPPF) module, which augments the standard SPPF with multi-scale spatial attention, channel attention preparation, and multi-head channel self-attention, and a hierarchical channel compression strategy that reduces the backbone width from 1024 to 512 and reduces C2f repetition counts. On RDD2022_China_Drone, the authors report a mAP50 of 67.6% with 0.89M parameters and 2.6 GFLOPs, a 2.11% relative improvement over YOLOv8n, and a claimed 16.8% improvement in the small-target D40 class. Generalization results on RDD2022_China_Motorbike are also reported. The experiments include comparisons with several YOLO variants, attention-based modifications, and prior road-damage detectors, plus an ablation over channel widths.

Significance. If the reported results are reproducible, the paper offers a useful efficiency-accuracy trade-off on a public benchmark, and the design of combining SPPF with bidirectional attention is a plausible direction for small-target road damage detection. Strengths include the use of a public dataset, comparison with a reasonable set of baselines, a separate generalization dataset, and a qualitative confusion-matrix analysis. However, the central attribution of the accuracy retention to BMS-SPPF is not directly supported by the ablation design, the headline per-class claim is not tabulated, and the single-run results do not establish the statistical significance of the reported 2.1% improvement. The manuscript is not circular: the claims are empirical and the method does not reduce to a fitted parameter.

major comments (3)
  1. [Section 5.3, Table 5] The ablation study varies the channel width and C2f repetition counts while simultaneously replacing SPPF with BMS-SPPF in every row. There is no row that keeps the standard SPPF at the compressed width (512) and the reduced repetition counts (2,3,3,2). Consequently, the difference between Ours_1024 and YOLO-ROC (69.1 to 67.6 mAP50) conflates two changes: the SPPF replacement and the compression scheme. The observed +2.9 point gain of Ours_1024 over YOLOv8n at full width does not by itself establish that BMS-SPPF provides the same gain at width 512, nor that the module is what compensates for compression. The paper should add an ablation row with standard SPPF at the compressed width and reduced C2f repeats; without it, the central claim in Sections 3.3 and 5.1 that BMS-SPPF compensates for information loss is not directly supported.
  2. [Abstract, Contributions (Section 1), Section 5] The claimed 16.8% improvement in mAP50 for the small-target D40 category is stated in the abstract and contributions but is never backed by a per-class results table or figure anywhere in the manuscript. The confusion matrix in Fig. 6 is not sufficient to recover per-class AP. Since the small-target improvement is a key mechanism of the proposed module, the authors must provide per-class AP for at least YOLOv8n and YOLO-ROC, and ideally for the ablation rows, so that the D40 claim can be independently verified.
  3. [Section 4.2 and Tables 2-5] All reported metrics come from single training runs with no seeds, no variance estimates, and no code release. The headline comparison is a 2.11% relative mAP50 improvement (66.2 to 67.6); in object detection benchmarks this margin is often within run-to-run variation, especially on a dataset of only 2,401 training images. The paper should report mean and standard deviation over at least three random seeds for the main comparison, the key ablations, and the D40 claim, and should state whether the differences are significant. Reproducibility would also be greatly improved by releasing code and training configuration files.
minor comments (5)
  1. [Section 3.3, Eq. (10) vs. Table 1] Equation (10) defines the repetition count as 2 for layers in {P3, P4} and 3 otherwise, which gives repeats (3,2,2,3) if applied to the four C2f stages, whereas Table 1 reports YOLOv8n repeats (3,6,6,3) changing to (2,3,3,2). Either the equation or the table is inconsistent, and the notation for stage indices should be clarified.
  2. [Section 3.3 and Section 5.1] The parameter reduction is reported as 67.3% in Section 3.3 and as 70.4% in the abstract and Section 5.1. These numbers should be reconciled; the correct value based on 3.01M to 0.89M is 70.4%.
  3. [Section 3.4, Eq. (11)] The loss function is described as a weighted sum of classification, localization, and objectness losses, but YOLOv8 is an anchor-free detector and does not use an objectness term. The authors should align the description with the actual YOLOv8 loss formulation or clarify how objectness is incorporated.
  4. [Section 5.2, Table 4] The abstract and Section 5.2 describe the generalization performance as 'excellent', but YOLO-ROC achieves 86.6% mAP50 on Motorbike compared with 86.9% for YOLOv8n and 89.0% for RT-DSAFDet; the claim should be softened to 'competitive' and the slight drop analyzed.
  5. [Throughout] There are numerous typographical and OCR-induced errors (e.g., 'liierarcliical', 'U.89l~I', 'CHiHa_DIOD€', 'haIulti-Head', 'less curve' for 'loss curve'). A careful proofreading pass is needed before publication.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity found; YOLO-ROC is a self-contained empirical benchmark with no equation-level reduction of outputs to inputs.

full rationale

The paper is an empirical object-detection study, not a derivation from first principles. The central claims (67.6% mAP50, 0.89M parameters, 2.6 GFLOPs, 2.0 MB) are measured benchmark results on the public RDD2022 dataset, and the accuracy comparisons in Table 2 are against external baselines with independent published results. No equation in the paper defines the reported mAP in terms of the proposed module, and no fitted parameter is renamed as a prediction. The cited self-work, RT-DSAFDet [23], is used only as a comparison target in Table 2 and Table 4, not as a premise that forces the paper's conclusions; the same applies to DynamicDet [14], which appears only in a related-work survey sentence. These self-citations are therefore not load-bearing and do not raise the circularity score. Two limitations flagged for correctness rather than circularity: (1) Table 5's ablation varies channel width and C2f repetition counts together with the SPPF replacement, and no row uses the standard SPPF at the compressed 512 width, so the isolated contribution of BMS-SPPF is not cleanly established; (2) the claimed 16.8% mAP50 improvement on the D40 small-target class is stated in the abstract and Contribution 1 but no per-class table or figure reports D40 results. Both are experimental-evidence gaps, not circular reductions of outputs to inputs. The loss function is explicitly the YOLOv8 baseline loss (Eq. 11), so there is no self-defined metric. The inference-speed and parameter-count results are independently measurable and not constructed from the claimed advantages. Overall, the manuscript is self-contained against external benchmarks and exhibits no significant circularity.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

All quantities in the ledger are architectural settings for the BMS-SPPF module and the compression scheme; the paper provides no formal derivation and no code, so the final model depends on these hand-set choices. The background assumptions are standard attention formulations and the fidelity of the public dataset and uncited baselines.

free parameters (5)
  • max_channel_width = 512
    Chosen from an ablation over 1024, 512, 256, and 128 in Table 5; the reported mAP depends on this choice.
  • C2f_repetition_counts = backbone (2,3,3,2), neck and head 2
    Architectural design choice in Section 3.3 and Table 1; no search procedure is reported.
  • MSSA_kernel_sizes = 3, 5, 7, 9
    Specified in Figure 2 caption and Section 3.2.2; chosen by hand, not justified experimentally.
  • MHSA_head_count = not stated
    Multi-head attention in Equation 8 requires a head count and head dimension, but these are not reported.
  • CAP_configuration = not stated (pooling vs lossless recombination, block size S)
    Section 3.2.3 describes two strategies but does not state which is used in the final model, nor the value of S in Equation 5.
assumptions (3)
  • domain assumption RDD2022 ground-truth annotations are correct and consistently labeled across D00, D10, D20, and D40.
    All mAP and confusion-matrix claims in Section 5 inherit the quality of the public RDD2022 labels; the paper does not audit or relabel them. Invoked in Section 4.1.
  • domain assumption The uncited baseline variants (SPPF_improve, SPPF_deformab, SPPF_uniRepLK, yolov8_MSFE, yolov8_ASCPA, yolov8_C2f_MS) were implemented faithfully under the same protocol.
    Tables 2 and 4 compare YOLO-ROC against these models, but no reference, code, or configuration is provided, so fairness of the comparison is assumed. Invoked in Section 4.4.
  • standard math Scaled dot-product multi-head self-attention and CBAM-style spatial and channel attention provide the feature-enhancement effects claimed in prior work.
    The paper imports standard attention formulations without re-deriving them; their benefit to small-target detection is treated as an established empirical result. Invoked in Sections 3.2.2 and 3.2.4.

how reviews work

0 comments
Cite this review

Pith. "Pith review of YOLO-ROC: A High-Precision and Ultra-Lightweight Model for Real-Time Road Damage Detection." pith.science (2026). https://pith.science/paper/IVQ7WK4K

@misc{pith2026250723225,
  author       = {Pith},
  title        = {Pith review of: YOLO-ROC: A High-Precision and Ultra-Lightweight Model for Real-Time Road Damage Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IVQ7WK4K}},
  note         = {Machine review of arXiv:2507.23225}
}
read the original abstract

Road damage detection is a critical task for ensuring traffic safety and maintaining infrastructure integrity. While deep learning-based detection methods are now widely adopted, they still face two core challenges: first, the inadequate multi-scale feature extraction capabilities of existing networks for diverse targets like cracks and potholes, leading to high miss rates for small-scale damage; and second, the substantial parameter counts and computational demands of mainstream models, which hinder their deployment for efficient, real-time detection in practical applications. To address these issues, this paper proposes a high-precision and lightweight model, YOLO - Road Orthogonal Compact (YOLO-ROC). We designed a Bidirectional Multi-scale Spatial Pyramid Pooling Fast (BMS-SPPF) module to enhance multi-scale feature extraction and implemented a hierarchical channel compression strategy to reduce computational complexity. The BMS-SPPF module leverages a bidirectional spatial-channel attention mechanism to improve the detection of small targets. Concurrently, the channel compression strategy reduces the parameter count from 3.01M to 0.89M and GFLOPs from 8.1 to 2.6. Experiments on the RDD2022_China_Drone dataset demonstrate that YOLO-ROC achieves a mAP50 of 67.6%, surpassing the baseline YOLOv8n by 2.11%. Notably, the mAP50 for the small-target D40 category improved by 16.8%, and the final model size is only 2.0 MB. Furthermore, the model exhibits excellent generalization performance on the RDD2022_China_Motorbike dataset.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

22 extracted references · 14 canonical work pages

  1. [1]

    Youwai, S., Chaiyaphat, A., Chaipetch, P.: YOLO9tr: a lightweight model for pavement damage detection utilizing a generalized efficient layer aggregation network and atten- tion mechanism. J. Real-Time Image Proc. 21, 163 (2024). https://doi.org/10.1007/511554-024-01545-2

    work page doi:10.1007/511554-024-01545-2 2024

  2. [2]

    Kulambayev, B., et al.: A Deep Learning-Based Approach for Road Surface Damage Detection. Con put. Mater. Contin. 73, 3403-3418 (2022). https://doi.org/10.32604/cmc.2022.029544

    work page arXiv 2022

  3. [3]

    Infrastructures

    Soilan, M., Sanchez Rodriguez, A., del Rio Barral, P., Pérez-Collazo, C., Arias, P., Riveiro, B.: Review of Laser Scanning Technologies and Their Applications for Road and Railway Infrastructure Monitoring. Infrastructures. 4, 58 (2019). https: //doi.org/ l0.3390/infrastructures4040058

    work page 2019

  4. [4]

    In: 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp

    Chen, Z., Lai, Z., Chen, J., Li, J.: Mind marginal non-crack regions: Clustering-inspired representation learning for crack segmentation. In: 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 12698-12708 (2024). https://doi.org/10.1109/CVPR52733.2024.01207

    arXiv 2024

  5. [5]

    In: Proceedings of the IEEE Conference on Computer Vi- sion and Pattern Recognition (CVPR), pp

    Redmond, J., Divvala, S., Girshick, R., Farhadi, A.: You Only Look Once: Unified, Real-Time Object Detection. In: Proceedings of the IEEE Conference on Computer Vi- sion and Pattern Recognition (CVPR), pp. 779-788 (2016). https://doi.org/10.1109/CVPR.20I6.9l

    work page doi:10.1109/cvpr.20i6.9l 2016

  6. [6]

    In: Leonardis, A., et al

    Wang, C.-Y., Yeh, I.-H., Liao, H.-Y.M.: YOLOv9: Learning What You Want to Learn Using Programmable Gradient Information. In: Leonardis, A., et al. (eds.) Computer Vision - ECCV 2024, pp. 1-21. Springer Nature Switzerland, Cham (2025). https://doi.org/I0.1007/978-3-031-72751-1_1

    work page 2025

  7. [7]

    Lan, X., Liu, L., Wang, X.: Dal-yolo: a multi-target de- tection model for UAV-based road maintenance integrating feature pyramid and attention mechanisms. J. Real-Time Image Proc. 22, 105 (2025). https://doi.org/10.I007/sI1554- 025-01685-z

    work page 2025

  8. [8]

    Zhao, M., Su, Y., Wang, J., et al.: MED-YOLOv8s: a new real-time road crack, pothole, and patch detec- tion model. J. Real-Time Image Proc. 21, 26 (2024) . https://doi.org/10.1007/sl1554-023-01405-5

    work page doi:10.1007/sl1554-023-01405-5 2024

Show all 22 references
  1. [9]

    In: Leonardis, A., et al

    Qin, D., et al.: MobileNetV4: Universal Models for the Mo- bile Ecosystem. In: Leonardis, A., et al. (eds.) Computer Vi- sion - ECCV 2024, pp. 78-96. Springer Nature Switzerland, Cham (2025). https://doi.org/10.1007/978-3-031-73661-2_5

    2025 doi

  2. [10]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp

    Zhang, X., Zhou, X., Lin, M., Sun, J.: Shuf- fleNet: An Extremely EH'1cient Convolutional Neu- ral Network for Mobile Devices. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 6848-6856 (2018). https://doi.org/10.1109/CVPR.2018.00716

    2018

  3. [11]

    IEEE Trans

    He, K., Zhang, X., Ren, S., Sun, J.: Spatial Pyramid Pool- ing in Deep Couvolutional Networks for Visual Recognition. IEEE Trans. Pattern Anal. Mach. If tell. 37(9), 1904-1916 (2015). https://doi.org/10.1109/TPAMI.2015.2389824

    2015

  4. [12]

    arXiv preprint arXiv:2408.15857 (2024)

    Yaseen, M.: What is YOLOv8: An In-Depth Exploration of the Internal Features of the Next-Generation Object Detector. arXiv preprint arXiv:2408.15857 (2024)

    2024 arXiv

  5. [13]

    arXiv preprint arXiv:2405.14458 (2024)

    Wang, A., et al.: YOLOv10: Real-Time End-to-End Ob- ject Detection. arXiv preprint arXiv:2405.14458 (2024)

    2024 arXiv

  6. [15]

    In: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp

    Zhang, G., Luo, Z., Tian, Z., Zhang, J., Zhang, X., Lu, S.: Towards Efficient Use of Multi-Scale Features in Transformer-Based Object Detectors. In: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 6206-6216 (2023). https://doi.or...

    2023

  7. [16]

    Park, J., Lee, J.-Y., Kweon, I.S.: CBAM: Con- volutional Block Attention Module

    Woo, S.. Park, J., Lee, J.-Y., Kweon, I.S.: CBAM: Con- volutional Block Attention Module. In: Ferrari, V., Hebert, M., Sminchisescu, C., Weiss, Y. (eds.) Computer Vision ECCV 2018, pp. 3-19. Springer International Publishing, Cham (2018). https://doi.org/10.1007/978-3-030-01234-2_1

    2018 doi

  8. [17]

    Han, T., Hou, S., Gao, C., et al.: EF-RT-DETR: a ef- Hcient focused real-time DETR model for pavement dis- tress detection. J. Real-Time Image Proc. 22, 63 (2025). https://doi.org/10.1007/sl1554-025-01641-x

    2025 doi

  9. [18]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp

    Lin, T.-Y., Dollar, P., Girshick, R., He, K., Hariharan, B., Belonging, S.: Feature Pyramid Networks for Object Detec- tion. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 936-944 (2017) . https://doi.org/10.1109/CVPR.2017.106

    2017 doi

  10. [19]

    In: Proceedings of the AAAI Conference on Artificial Intelligence, vol

    Zheng, Z., Wang, P., Liu, W., Li, J., Ye, R., Ren, D.: Distance-IoU Loss: Faster and Better Learning for Bounding Box Regression. In: Proceedings of the AAAI Conference on Artificial Intelligence, vol. 34, pp. 12993-13000 (2020). https://doi.org/10.1609/aaai.v34i07.6999

    2020 doi

  11. [20]

    arXiv preprint arXiv:2209.08538 (2022)

    Arya, D., Maeda, H., Ghosh, S., Toshniwal, D., Sekimoto, Y.: RDD2022: A multi-national image dataset for automatic Road Damage Detection. arXiv preprint arXiv:2209.08538 (2022)

    2022 arXiv

  12. [21]

    Tsirtsakis, P., Zacharis, G., Maraslidis, G.S., Frag- ulis, G.F.: Deep learning for object recognition: A comprehensive review of models and algorithms. Int. J. Cogn. Con put. Eng. 6, 298-312 (2025). https://doi.org/10.1016/j.ijcce.2025.0l.004

    2025 doi

  13. [22]

    la/lultimed

    Chen, W., Luo, J., Zhang, F., Tian, Z.: A re- view of object detection: Datasets, performance eval- uation, architecture, applications and current trends. la/lultimed. Tools Appl. 83(24), 6560365661 (2024). https://doi.org/10.1007/sl1042-023-17949-4

    2024 doi

  14. [23]

    Pan, W., Wang, X., Huan, W.: Real-time dynamic scale- aware fusion detection network: take road damage detection as an example. J. Real-Time Image Proc. 22, 55 (2025). https://doi.org/10.l007/sl1554-025-01634-w

    2025

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.