pith. sign in

arxiv: 2606.12066 · v1 · pith:ZJQAZDCCnew · submitted 2026-06-10 · 💻 cs.CV

Performance Analysis of YOLOv11 and YOLOv8 for Mixed Traffic Object Detection under Adverse Weather Conditions in Developing Countries

Quoc Thuan Nguyen , Ha Anh Vu , Ngo Dang Thanh Ngan , Minh Phuc Hoang Ngoc This is my paper

Pith reviewed 2026-06-27 10:20 UTC · model grok-4.3

classification 💻 cs.CV
keywords YOLOv11YOLOv8object detectionadverse weathermixed trafficautonomous drivingedge deploymentprecision
0
0 comments X

The pith

YOLOv11 Nano reaches 46.6% mAP@50 and 3.2% higher precision than YOLOv8 Nano on fused driving data in rain and low light.

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

The paper benchmarks the YOLOv11 Nano model against YOLOv8 Nano on a dataset created by combining the Indian Driving Dataset and Berkeley Deep Drive Dataset. It reports that YOLOv11n delivers 46.6% mean average precision at the 50% threshold, a 3.2% precision gain that cuts false positives in dense traffic, and 22% lower floating-point operations while running at 70.9 frames per second on a Tesla T4 GPU. These results address object detection needs for autonomous systems in developing countries where mixed traffic, rain, and poor lighting are common. The evaluation focuses on accuracy, speed, and energy use together rather than any single metric.

Core claim

YOLOv11n achieves a mean Average Precision (mAP@50) of 46.6% with a 3.2% improvement in Precision over YOLOv8n on the custom fused dataset of dense mixed traffic under rain and low-light conditions, while requiring 22% fewer FLOPs (6.3G versus 8.1G) at a real-time speed of 70.9 FPS on a Tesla T4 GPU.

What carries the argument

Direct comparison of YOLOv11 Nano and YOLOv8 Nano architectures on the fused IDD plus BDD100K dataset for detection under adverse weather.

Load-bearing premise

The measured gains come from architectural differences in YOLOv11 rather than from how the two models were trained or how the source datasets were combined and balanced.

What would settle it

Retrain both YOLOv11n and YOLOv8n from identical random initializations on the exact same fused dataset split using the same augmentation and optimization schedule, then check whether the 3.2% precision and 22% FLOP gaps persist.

read the original abstract

In modern vehicular systems, robust performance under harsh conditions has become a critical problem of autonomous driving. Our study delivers a comprehensive evaluation of the newest iteration of the YOLO series, which is YOLOv11 Nano architecture benchmarked against the widely adopted YOLOv8 Nano as a baseline on a custom fused dataset that combines the Indian Driving Dataset (IDD) [1] and Berkeley Deep Drive Dataset (BDD100K) [2]. We have analyzed the trade-offs among detection accuracy, inference speed, and computational efficiency in high-entropy scenarios involving dense mixed traffic, rain, and low-light conditions. Specifically, YOLOv11n achieves a mean Average Precision (mAP@50) of 46.6%, with a notable 3.2% improvement in Precision over the baseline, effectively reducing false positives in cluttered scenes. Furthermore, the proposed model exhibits enhanced energy efficiency, requiring 22% fewer FLOPs (6.3G vs. 8.1G) while maintaining real-time inference speed of 70.9 FPS on a Tesla T4 GPU, offering an optimal trade-off for safety-critical edge deployment.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The manuscript evaluates YOLOv11 Nano against YOLOv8 Nano on a fused IDD+BDD100K dataset for object detection in dense mixed traffic under rain and low-light conditions typical of developing countries. It claims YOLOv11n delivers 46.6% mAP@50, a 3.2% precision gain over the baseline (reducing false positives), 22% fewer FLOPs (6.3G vs. 8.1G), and 70.9 FPS on a Tesla T4 GPU, positioning it as an efficient choice for safety-critical edge deployment.

Significance. If the numeric gains are shown to result from architectural differences under controlled identical training conditions and are reproducible, the work would supply useful empirical guidance for selecting lightweight detectors suited to resource-limited autonomous-driving scenarios in high-entropy traffic environments. The emphasis on adverse weather and developing-country datasets addresses a practically relevant gap, though current unverifiability restricts immediate adoption.

major comments (2)
  1. [Abstract] Abstract: The reported performance deltas (46.6% mAP@50, 3.2% precision lift, 22% FLOP reduction) are presented as direct outcomes of YOLOv11n versus YOLOv8n, yet no information is supplied on training protocol, loss functions, hyperparameter search, optimizer settings, learning-rate schedules, data-augmentation pipelines, or the exact fusion and class-balancing procedure applied to IDD+BDD100K. This omission renders the central attribution of gains to architecture unverifiable.
  2. [Experimental setup] Experimental setup (throughout manuscript): No statement confirms that both models received identical training schedules, augmentation, or balancing weights during dataset fusion. Without this control, the observed differences cannot be isolated from potential confounding variables, directly undermining the claim that YOLOv11n provides an "optimal trade-off" due to its design.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments regarding the need for detailed experimental information. We will revise the manuscript to provide comprehensive details on training protocols and explicitly confirm identical conditions for both models.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The reported performance deltas (46.6% mAP@50, 3.2% precision lift, 22% FLOP reduction) are presented as direct outcomes of YOLOv11n versus YOLOv8n, yet no information is supplied on training protocol, loss functions, hyperparameter search, optimizer settings, learning-rate schedules, data-augmentation pipelines, or the exact fusion and class-balancing procedure applied to IDD+BDD100K. This omission renders the central attribution of gains to architecture unverifiable.

    Authors: We agree that the manuscript currently lacks explicit details on the training protocol, hyperparameters, loss functions, optimizer settings, learning-rate schedules, data-augmentation pipelines, and the fusion/class-balancing procedure for IDD+BDD100K. In the revised version, we will add a dedicated subsection in the Experimental Setup section that fully documents these elements, including the exact values and procedures used. This addition will allow verification that performance differences arise under controlled, identical conditions. revision: yes

  2. Referee: [Experimental setup] Experimental setup (throughout manuscript): No statement confirms that both models received identical training schedules, augmentation, or balancing weights during dataset fusion. Without this control, the observed differences cannot be isolated from potential confounding variables, directly undermining the claim that YOLOv11n provides an "optimal trade-off" due to its design.

    Authors: We acknowledge that the current text does not explicitly confirm identical training conditions for YOLOv11n and YOLOv8n. The revised manuscript will include a clear statement in the Experimental Setup section affirming that both models were trained with the same schedules, augmentations, and balancing weights. We will also provide the specific hyperparameter values and procedures to isolate architectural effects. revision: yes

Circularity Check

0 steps flagged

No circularity: direct empirical benchmarking with no derivations or self-referential reductions

full rationale

The paper is a straightforward empirical comparison of YOLOv11n vs YOLOv8n on a fused IDD+BDD100K dataset under adverse conditions. It reports measured metrics (mAP@50 = 46.6%, precision lift of 3.2%, FLOPs 6.3G vs 8.1G, 70.9 FPS) as experimental outcomes without any equations, fitted parameters, predictions derived from inputs, or load-bearing self-citations. Dataset citations [1][2] are external and non-overlapping. No step reduces by construction to the paper's own inputs; the central claims remain falsifiable measurements.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The evaluation rests on the unstated premise that the two models were trained and evaluated under identical conditions on a representative fused dataset; no free parameters, new axioms, or invented entities are introduced in the abstract.

axioms (1)
  • domain assumption The two YOLO variants were trained and evaluated under comparable conditions on the fused dataset.
    Required for attributing observed differences to model version rather than experimental setup.

pith-pipeline@v0.9.1-grok · 5755 in / 1372 out tokens · 40228 ms · 2026-06-27T10:20:51.257982+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

19 extracted references · 2 canonical work pages

  1. [1]

    In: 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), pp

    Varma, G., Subramanian, A., Namboodiri, A., Chandraker, M., Jawahar, C.V.: IDD: A Da- taset for Exploring Problems of Autonomous Navigation in Unconstrained Environments. In: 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), pp. 1743– 1751 (2019)

    2019

  2. [2]

    In: CVPR, pp

    Yu, F., Chen, H., Wang, X., Xian, W., Chen, Y., Liu, F., Madhavan, V., Darrell, T.: BDD100K: A Diverse Driving Dataset for Heterogeneous Multitask Learning. In: CVPR, pp. 2636–2645 (2020)

    2020

  3. [3]

    arXiv preprint arXiv:2410.07701 (2024)

    Min, C., Si, S., Wang, X., et al.: Autonomous Driving in Unstructured Environments: How Far Have We Come. arXiv preprint arXiv:2410.07701 (2024)

    arXiv 2024

  4. [4]

    IEEE Access 11, 134877 – 134890 (2023)

    Wibowo, A., Trilaksono, B.R., Hidayat, E.M.I., Munir, R.: Object Detection in Dense and Mixed Traffic for Autonomous Vehicles With Modified Yolo. IEEE Access 11, 134877 – 134890 (2023)

    2023

  5. [5]

    IEEE Ac- cess 13, 93711–93725 (2025)

    Wei, J., As’arry, A., Rezali, K.A.M., Yusoff, M.Z.M., Ma, H., Zhang, K.: A Review of YOLO Algorithm and Its Applications in Autonomous Driving Object Detection. IEEE Ac- cess 13, 93711–93725 (2025)

    2025

  6. [6]

    Computers 13, 336 (2024)

    Ali, M.L., Zhang, Z.: The YOLO Framework: A Comprehensive Review of Evolution, Ap- plications, and Benchmarks in Object Detection. Computers 13, 336 (2024). 12

    2024

  7. [7]

    In: 2024 7th International Confer- ence on Informatics and Computational Sciences (ICICoS)

    Saha, U., Ahamed, I.U., Hossain, M.I.: YOLOv8 for Bangla license plate recognition: ad- vancing real-time object detection in localized contexts. In: 2024 7th International Confer- ence on Informatics and Computational Sciences (ICICoS). IEEE (2024). https://doi.org/10.1109/ICICoS62600.2024.10636876. Last accessed 2026/01/29

    work page doi:10.1109/icicos62600.2024.10636876 2024

  8. [8]

    Procedia Computer Science 246, 2792–2801 (2024)

    Alahdala, N.M., Abukhodair, F., Meftah, L.H., Cherif, A.: Real -time Object Detection in Autonomous Vehicles with YOLO. Procedia Computer Science 246, 2792–2801 (2024)

    2024

  9. [9]

    Plants 14(5), 653 (2025)

    Lee, Y.-S., Patil, M., Kim, J., Seo, Y.B., Ahn, D., Kim, G.-D.: Hyperparameter optimization for tomato leaf disease recognition based on YOLOv11m. Plants 14(5), 653 (2025). https://doi.org/10.3390/plants14050653

    work page doi:10.3390/plants14050653 2025

  10. [10]

    In: Wallach, H

    Paszke, A., et al.: PyTorch: An Imperative Style, High-Performance Deep Learning Library. In: Wallach, H. et al. (eds.) Advances in Neural Information Processing Systems, vol. 32. Curran Associates, Inc. (2019)

    2019

  11. [11]

    https://github.com/ultralyt- ics/ultralytics

    Jocher, G., Chaurasia, A., Qiu, J.: Ultralytics YOLO (2023). https://github.com/ultralyt- ics/ultralytics. Last accessed 2026/01/28

    2023

  12. [12]

    In: Fleet, D., Pajdla, T., Schiele, B., Tuytelaars, T

    Lin, T.-Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, C.L.: Microsoft COCO: Common Objects in Context. In: Fleet, D., Pajdla, T., Schiele, B., Tuytelaars, T. (eds.) ECCV 2014, LNCS, vol. 8693, pp. 740–755. Springer, Cham (2014)

    2014

  13. [13]

    https://docs.ultralyt- ics.com/guides/yolo-performance-metrics/

    Ultralytics: YOLO Performance Metrics - Ultralytics YOLOv8 Docs. https://docs.ultralyt- ics.com/guides/yolo-performance-metrics/. Last accessed 2026/01/28

    2026

  14. [14]

    In: Proceedings of the European Conference on Computer Vision (ECCV)

    Ma, N., Zhang, X., Zheng, H.-T., Sun, J.: ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 116–131 (2018)

    2018

  15. [15]

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

    Ding, X., Zhang, X., Ma, N., Han, J., Ding, G., Sun, J.: RepVGG: Making VGG-style Con- vNets Great Again. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13733–13742 (2021)

    2021

  16. [16]

    In: Proceedings of the AAAI Conference on Ar- tificial Intelligence, vol

    Liu, W., Ren, G., Yu, R., Guo, S., Zhu, J., Zhang, L.: Image -Adaptive YOLO for Object Detection in Adverse Weather Conditions. In: Proceedings of the AAAI Conference on Ar- tificial Intelligence, vol. 36(2), pp. 1792–1800 (2022)

    2022

  17. [17]

    IEEE Signal Processing Magazine 38(1), 55–70 (2021)

    Hnewa, M., Radha, H.: Object Detection Under Rainy Conditions for Autonomous Vehicles: A Review of State-of-the-Art and Emerging Techniques. IEEE Signal Processing Magazine 38(1), 55–70 (2021)

    2021

  18. [18]

    arXiv preprint arXiv:2507.19912 (2025)

    Kumar, R., Reddy, D.S., Rajalakshmi, P.: DriveIndia: An Object Detection Dataset for Di- verse Indian Traffic Scenes. arXiv preprint arXiv:2507.19912 (2025)

    arXiv 2025

  19. [19]

    Sensors 24(19), 6319 (2024)

    Padia, A., et al.: Object Detection and Classification Framework for Analysis of Video Data Acquired from Indian Roads. Sensors 24(19), 6319 (2024)

    2024