Optimizing Helmet Detection with Hybrid YOLO Pipelines: A Detailed Analysis

Balamurali S; Dejey D; Vaikunth M; Vishaal C

REVIEW 4 major objections 6 minor 20 references

Optimizing Helmet Detection with Hybrid YOLO Pipelines: A Detailed Analysis

T0 review · 4 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read This paper claims that inserting a small CNN feature-feeding block before YOLOv8, YOLOv9, and YOLOv11 improves helmet-detection precision, recall, and mAP@50 by 2-3 percentage points over the same models run alone.

desk verdict A routine YOLO-hybrid comparison whose only real claim is undermined by missing controls and a suspicious timing reversal. read the letter →

arxiv 2412.19467 v1 pith:ZF5ZSSSY submitted 2024-12-27 cs.CV cs.AIcs.LG

Vaikunth M , Dejey D , Vishaal C , Balamurali S This is my paper · ORCID

classification cs.CVcs.AIcs.LG

keywords helmetdetectionYOLOv8YOLOv9YOLOv11hybridCNN-YOLOpipelineobjectmAP@50trafficsafety

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

The paper is trying to establish that helmet detection in traffic imagery is better served by a hybrid pipeline than by a plain YOLO detector on its own. It compares YOLOv8, YOLOv9, and YOLOv11 with and without a lightweight CNN stage placed before the detector, and reports that the hybrid versions beat the plain versions by 2-3 percentage points in mAP@50 on a dataset of over 3,500 helmet images. If true, the result matters because automated helmet enforcement needs a detector that is reliable enough for traffic cameras while still running in real time. The paper also argues that h-YOLOv11 is the practical choice: its accuracy is nearly as high as h-YOLOv9's, and its inference time is far lower.

What carries the argument

The central object is the hybridization block: a lightweight sequential CNN whose layers apply convolution, batch normalization, and activation, with only filter counts, dimensions, and padding varying between layers. This block sits before the YOLO model and feeds extracted features into the detector; the paper states that it is the only architectural difference between each h-YOLO model and its plain counterpart, which is what makes the paired comparison meaningful.

What would settle it

Retrain the three YOLO versions with and without the CNN block under a single published hyperparameter configuration, identical augmentation, and the same data split, then evaluate on a held-out helmet set; if the mAP@50 gap falls below the reported 2-3% or changes sign, the central claim is refuted.

Watch

Extended reading notes

Core claim

On the authors' helmet-detection dataset, adding a three-block lightweight CNN before the YOLO detector raises precision, recall, and mAP@50 for every YOLO version tested: h-YOLOv8 reaches 0.925 mAP@50 versus 0.905, h-YOLOv9 reaches 0.932 versus 0.906, and h-YOLOv11 reaches 0.914 versus 0.892. The authors conclude that h-YOLO models are preferable to independent YOLO models for this safety task because a 2-3% accuracy gain is significant when the system may affect civilian life, and that h-YOLOv11 provides the best balance of detection quality and speed.

Load-bearing premise

The comparison is only fair if the plain and hybrid runs used identical data preprocessing, augmentation, hyperparameters, and dataset splits, with the added CNN block as the sole difference; the paper asserts this but reports no settings or split details to verify it.

Editorial extensions

If this is right

If a traffic camera runs any of the hybrid detectors, it should expect roughly 2-3 percentage points higher mAP@50 than the corresponding plain YOLO model, at the cost of only a few extra minutes of training time.
For real-time deployment, h-YOLOv11 is the recommended model because its mAP trails h-YOLOv9 by about 1.8 points while its testing time drops from 36.4 ms to 11.9 ms.
The same hybrid recipe can be applied to detect other rider safety equipment, such as rear-view mirrors or riding shoes, which the paper names as future work.
The measured testing times place the hybrid detectors in the range needed for live traffic-camera screening rather than only offline analysis.
Because the accuracy gain is reported consistently across three YOLO generations, the benefit appears to come from the added CNN stage itself rather than from any single YOLO version.

Reading between the lines

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

The paper does not test generalization outside its training distribution; a fair next experiment would evaluate the hybrid and plain models on a separate, independently collected helmet dataset to see whether the 2-3% gap holds under domain shift.
A natural extension the authors leave implicit is threshold tuning for the enforcement use case, where false negatives, missed riders without helmets, may be costlier than false positives; recall-oriented operating points would show whether the hybrid advantage persists where it matters most.
The hybrid block's benefit might transfer to other single-stage detectors or newer YOLO releases, but the paper provides no evidence for that transfer, so that remains an open hypothesis.
The claimed real-time suitability is based on average inference time per image; stress tests with multiple simultaneous camera feeds or low-power edge hardware would be needed before accepting the deployment claim.

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.

Desk Editor's Note

read the letter

The one thing to know: this is a small applied study claiming that a lightweight CNN front-end adds 2-3% mAP to YOLOv8/9/11 on helmet detection, but the comparison is a single run with no controls, so the claim is not established. The stress-test note is right, and it holds up on reading.

What's actually new: the specific h-YOLO variants and their benchmark numbers on this helmet dataset. That is it. The components are standard YOLO models plus a conventional CNN feature-feeding block; the paper itself cites prior ensembling and YOLO modification work. Credit where due: the paper is readable, the tables are internally consistent, and the plain YOLO baselines behave as expected. The authors also made a sensible choice to report mAP@50, precision, and recall, though they then fail to provide variance estimates.

The soft spots are load-bearing. Section 3.6 asserts that the independent and hybrid pipelines are identical except for the hybridization block, but no hyperparameters, augmentation details, or train/test split sizes or seeds are given. On a ~3500-image dataset, a 2-3% mAP gap can easily be split noise. There are no repeated runs, no confidence intervals, no significance tests. The paper's phrase 'remarkably improves' is not supported by the evidence as reported.

The most telling anomaly is in Table 3: h-YOLOv9 has a testing time of 36.4 ms while plain YOLOv9 is 43.0 ms, despite the extra CNN block. That is hard to reconcile with the 'only varying parameter' claim and suggests different hardware, batching, or other conditions. The authors don't explain this. Also, the description of YOLOv9 in Section 3.6.2 is factually wrong — YOLOv9 does not use a hybrid CNN-Transformer backbone; it uses GELAN and PGI. The literature survey is mostly descriptive but contains this error, which makes me cautious about the rest.

The conclusion overreaches: the paper says the 2-3% difference is 'significant considering the objective... can potentially deal with civilian life,' but 'significant' is doing a lot of work. The authors never define a threshold for significance, and the evidence doesn't support it.

Who is this for? Someone working on helmet detection for traffic enforcement might find the method section a useful starting point, but they would need to redo the experiments properly. This is not a serious-referee paper as it stands. If the authors added repeated runs, confidence intervals, full hyperparameter disclosure, and released code and data, it could become a modest empirical note. As is, I would not send it out for review; a desk reject with encouragement to resubmit after a proper controlled study is the right call.

Referee Report

4 major / 6 minor

Summary. The paper proposes a hybrid object-detection pipeline (h-YOLO) that prepends a lightweight custom CNN feature-extraction block to YOLOv8, YOLOv9, and YOLOv11, and compares these hybrid models against their plain counterparts for motorcycle-helmet detection. Using a dataset of over 3500 images, the authors report precision, recall, mAP@50, and training/testing times. They claim that the h-YOLO variants outperform the plain YOLO models by about 2-3% mAP@50, that h-YOLOv9 has the best accuracy, and that h-YOLOv11 offers the best speed-accuracy trade-off. The manuscript includes a literature survey, a description of the proposed architecture, evaluation metrics, and a discussion of results.

Significance. If the central claim were established, the paper would provide a simple, practical recipe for improving helmet-detection accuracy in traffic-safety applications by augmenting recent YOLO models with a lightweight CNN front-end. The task is socially relevant and the comparison across YOLOv8, YOLOv9, and YOLOv11 is timely. The paper also reports both accuracy and timing metrics, which is useful for real-time deployment considerations. However, the empirical evidence is currently weakened by the absence of controlled experimental details, variance estimates, and significance testing; the contribution is therefore not yet at the level required for a journal publication.

major comments (4)

[§3.6 and §4] The load-bearing claim that the hybridization block is the 'only varying parameter' between the h-YOLO and plain YOLO models is unverifiable because the manuscript reports no training hyperparameters (epochs, batch size, learning rate, optimizer scheduling), no augmentation configuration, no train/validation/test split ratio, and no random seeds. Without this information, the 2-3% mAP@50 differences in Table 2 cannot be attributed to the hybridization block rather than to uncontrolled differences in the training setup. Please report the full experimental configuration for all six models.
[Table 2 and §4] Each model is evaluated with a single point estimate per metric, with no repeated runs, confidence intervals, or significance tests. On a dataset of roughly 3500 images (§3.1), an absolute mAP@50 difference of 0.02-0.03 (e.g., 0.925 vs 0.905 for v8; 0.932 vs 0.906 for v9) can easily fall within run-to-run variance. Therefore the statement 'the accuracy difference is 2-3% which is significant considering the objective of the project' is not supported as written. Provide variance estimates or statistical tests, or temper the claim.
[Table 3] The reported test time for h-YOLOv9 (36.4 ms) is lower than that for plain YOLOv9 (43.0 ms), even though the h-YOLO pipeline adds a CNN block before the same YOLO model. This is inconsistent with the claim that the only difference between the paired models is the hybridization block. It suggests that the inference-time measurements were made under different conditions (e.g., hardware, batch size, or number of averaged runs). Please clarify the measurement setup and re-measure under controlled conditions.
[§3.3 and Figure 2] The architecture of the proposed CNN hybridization block is under-specified. The text describes the first CNN layer and says the subsequent layers differ in 'number of filters, dimensions, and padding,' but the actual filter counts, kernel sizes, padding values, and number of layers are not given, and no code or pretrained weights are provided. This makes the proposed method irreproducible. Please include a complete architectural specification or release the implementation.

minor comments (6)

[§4, first paragraph] The text says 'From Table 1, it is understood that the h-YOLO models categorically produce better precision, recall, and mAP' and later 'as shown in Table 2' for training and testing times; these references are wrong: performance metrics are in Table 2, and training/testing times are in Table 3.
[§3.7, equation for recall] The word 'Recall' is misspelled as 'Recal' in the equation.
[Table 1, reference [8]] Reference [8] lists 'T.-Y. Ross and G. Dollár' as the authors of RetinaNet; the correct first author is T.-Y. Lin (Tsung-Yi Lin).
[§3.6.2] The statement that YOLOv9 is 'the first in the series to utilize a hybrid CNN-Transformer backbone' is inaccurate; YOLOv9's architecture uses GELAN and PGI and does not feature a transformer backbone. Please verify and correct.
[§3.4] Hyperparameter tuning is described qualitatively, but no actual values for learning rate, batch size, number of epochs, or optimizer settings are reported anywhere in the manuscript; adding a table of hyperparameters would aid reproducibility.
[§3.5] The mention of 'image fusion' for refining detection accuracy is vague and not used in the reported experiments; either explain the technique and its role or remove it.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper is an empirical benchmark comparison with no fitted-parameter or self-citation chain; concerns are experimental control and reproducibility, not circular reasoning.

full rationale

The paper's central claim is an empirical comparison between h-YOLO and plain YOLO models using precision, recall, and mAP@50 (Tables 2 and 3). There is no analytical derivation in which an output is defined in terms of an input, no fitted constant later relabeled as a prediction, and no load-bearing uniqueness theorem. The only self-citation, 'M. Vaikunth et al. [This work]' in Table 1, is a placeholder for the present work and carries no evidential weight. Sections 3.3 and 3.6 describe a fixed data pipeline and assert that the hybridization block is the only varying parameter, but the paper does not report hyperparameters, augmentation details, splits, seeds, or repeated runs; this is a reproducibility and validity limitation, not circularity. The conclusion that 'the h-YOLO models are preferred' because a 2-3% mAP difference is 'significant' rests on an uncontrolled single-run comparison, but that concern belongs to soundness of the empirical claim rather than to the derivation being equivalent to its inputs. The paper is self-contained against external benchmarks (YOLO baselines trained on the same dataset), so no circular step is exhibited.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on an empirical comparison; no new physical entities are introduced. The main burdens are unreported hyperparameters, an unreleased dataset, and the controlled-comparison assumption.

free parameters (3)

Training hyperparameters (epochs, batch size, learning rate, optimizer schedule) = not reported
Section 3.4 says these were 'carefully tuned' but gives no values, so the reported results depend on unreported choices.
CNN layer dimensions, filter counts, and padding = not reported
Section 3.3 says the three CNN layers differ in filters, dimensions, and padding, but none are specified, so the hybrid architecture is underdetermined.
Train/test split ratio = not reported
Section 3.3 states data is split into training and testing sets, but the ratio and stratification are not given.

assumptions (4)

domain assumption The combined Roboflow and custom phone-image dataset is representative of real-world helmet detection and its labels are correct.
The entire evaluation rests on this dataset (Section 3.1); no label audit or class distribution is reported.
domain assumption The only difference between h-YOLO and plain YOLO is the CNN hybridization block.
This controlled-comparison assumption (Section 3.6) is needed to attribute the 2-3% mAP gain to the CNN block, but implementation consistency is not verifiable from the text.
domain assumption The reported precision, recall, and mAP@50 numbers are computed correctly with a fixed IoU threshold of 50%.
Section 3.7 defines the metrics, but no evaluation code or raw predictions are provided to check the numbers.
domain assumption YOLOv11 has fewer parameters and faster training/inference than YOLOv8/v9, as stated in official documentation.
Used in Section 4 to explain h-YOLOv11's speed; the claim is not independently verified in the paper.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Optimizing Helmet Detection with Hybrid YOLO Pipelines: A Detailed Analysis." pith.science (2026). https://pith.science/paper/ZF5ZSSSY

@misc{pith2026241219467,
  author       = {Pith},
  title        = {Pith review of: Optimizing Helmet Detection with Hybrid YOLO Pipelines: A Detailed Analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZF5ZSSSY}},
  note         = {Machine review of arXiv:2412.19467}
}

read the original abstract

Helmet detection is crucial for advancing protection levels in public road traffic dynamics. This problem statement translates to an object detection task. Therefore, this paper compares recent You Only Look Once (YOLO) models in the context of helmet detection in terms of reliability and computational load. Specifically, YOLOv8, YOLOv9, and the newly released YOLOv11 have been used. Besides, a modified architectural pipeline that remarkably improves the overall performance has been proposed in this manuscript. This hybridized YOLO model (h-YOLO) has been pitted against the independent models for analysis that proves h-YOLO is preferable for helmet detection over plain YOLO models. The models were tested using a range of standard object detection benchmarks such as recall, precision, and mAP (Mean Average Precision). In addition, training and testing times were recorded to provide the overall scope of the models in a real-time detection scenario.

Figures

Figures reproduced from arXiv: 2412.19467 by the authors.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

20 extracted references · 19 canonical work pages

[1]

You only look once: Unified, real -time object detection,

J. Redmon, “You only look once: Unified, real -time object detection,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2016

work page 2016
[2]

A comprehensive review of yolo architectures in computer vision: From yolov1 to yolov8 and yolo -nas,

J. Terven, D. -M. Córdova-Esparza, and J. -A. Romero-González, “A comprehensive review of yolo architectures in computer vision: From yolov1 to yolov8 and yolo -nas,” Machine Learning and Knowledge Extraction, vol. 5, no. 4, pp. 1680–1716, 2023

work page 2023
[3]

Complexer-yolo: Real-time 3d object detection and tracking on semantic point clouds ,

M. Simon, K. Amende, A. Kraus, J. Honer, T. Samann, H. Kaulbersch, S. Milz, and H. Michael Gross, “Complexer-yolo: Real-time 3d object detection and tracking on semantic point clouds ,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2019, pp. 0–0

work page 2019
[4]

Object detection and tracking with yolo and the sliding innovation filter,

A. Moksyakov, Y. Wu, S. A. Gadsden, J. Yawney, and M. AlShabi, “Object detection and tracking with yolo and the sliding innovation filter,” Sensors, vol. 24, no. 7, p. 2107, 2024

work page 2024
[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, 2016

work page 2016
[6]

Improvement of object detection based on faster r -cnn and yolo,

J. Fan, J. Lee, I. Jung, and Y. Lee, “Improvement of object detection based on faster r -cnn and yolo,” in 2021 36th International Technical Conference on Circuits/Systems, Computers and Communications (ITC-CSCC). IEEE, 2021, pp. 1–4

work page 2021
[7]

SSD: Single shot multibox detector,

W. Liu, D. A nguelov, D. Erhan, C. Szegedy, S. Reed, C. -Y. Fu, and A. C. Berg, “SSD: Single shot multibox detector,” in Computer Vision –ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11–14, 2016, Proceedings, Part I 14. Springer, 2016, pp. 21–37

work page 2016
[8]

Focal loss for dense object detection,

T.-Y. Ross and G. Dollár, “Focal loss for dense object detection,” in proceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 2980–2988

work page 2017

Show all 20 references

[9]

Real -time object detection using an en semble of one stage and two stage object detection models with dynamic fine -tuning using kullback-leibler divergence,

L. Nkalubo, R. Nakibuule, and N. Okila, “Real -time object detection using an en semble of one stage and two stage object detection models with dynamic fine -tuning using kullback-leibler divergence,” Authorea Preprints, 2023. [Online]. Available: https://doi.org/10.22541/au.1...

2023
[10]

Mask r -cnn,

K. He, G. Gkioxari, P. Dollár, and R. Gi rshick, “Mask r -cnn,” in Proceedings of the IEEE international conference on computer vision, 2017, pp. 2961–2969

2017
[11]

Bike helmet detection dataset,

B. Helmets, “Bike helmet detection dataset,” https://universe.roboflow.com/bike -helmets/bike- helmet-detection-2vdjo, sep 2021, visited on 202 4-11-09. [Online]. Available: https://universe.roboflow.com/bike-helmets/bike-helmet-detection-2vdjo

2021
[12]

Training object detection and recognition cnn models using data augmentation,

D. M. Montserrat, Q. Lin, J. Allebach, and E. J. Delp, “Training object detection and recognition cnn models using data augmentation,” Electronic Imaging, vol. 29, pp. 27–36, 2017

2017
[13]

Dmac-yolo: A high-precision yolo v5s object detection model with a novel optimizer,

C. Meng, S. Liu, and Y. Yang, “Dmac-yolo: A high-precision yolo v5s object detection model with a novel optimizer,” in 2024 International Joint Conference on Neural Networks (IJCNN). IEEE, 2024, pp. 1–8

2024
[14]

Adam optimizer based deep learning approach for improving efficiency in license plate recognition,

P. K. Rajput, K. K. Ravu lakollu, N. Jagadam, and P. Singh, “Adam optimizer based deep learning approach for improving efficiency in license plate recognition,” in Automation and Computation. CRC Press, 2023, pp. 439–449

2023
[15]

Yolo -firi: Improved yolov5 for infrared image object detection,

S. Li, Y. Li, Y. Li, M. Li, and X. Xu, “Yolo -firi: Improved yolov5 for infrared image object detection,” IEEE access, vol. 9, pp. 141 861–141 875, 2021

2021
[16]

Sod -yolov8—enhancing yolov8 for small object detection in aerial imagery and traffic scenes,

B. Khalili and A. W. Smyth, “Sod -yolov8—enhancing yolov8 for small object detection in aerial imagery and traffic scenes,” Sensors, vol. 24, no. 19, p. 6209, 2024

2024
[17]

What is yolov9: An in -depth exploration of the internal features of the next -generation object detector,

M. Yaseen, “What is yolov9: An in -depth exploration of the internal features of the next -generation object detector,” arXiv preprint arXiv:2409.07813, 2024

2024 arXiv
[18]

Yolov11 for vehicle detection: Advancements, performance, and applications in intelligent transportation systems,

M. A. R. Alif, “Yolov11 for vehicle detection: Advancements, performance, and applications in intelligent transportation systems,” arXiv preprint arXiv:2410.22898, 2024

2024 arXiv
[19]

Efficient -lightweight yolo: Improving small object detection in yolo for aerial images,

M. Hu, Z. Li, J. Yu, X. Wan, H. Tan, and Z. Lin, “Efficient -lightweight yolo: Improving small object detection in yolo for aerial images,” Sensors, vol. 23, no. 14, p. 6423, 2023. 92 Computer Science & Information Technology (CS & IT) AUTHORS Vaikunth M is currently pursuing ...

2023
[2024]

Vishaal C is a final -year undergraduate student in Computer Science and Engineering at the College of Engineering, Guindy, Chennai

Her research interests include Signal, Image and Video Processing, IoT, and Multimedia Security. Vishaal C is a final -year undergraduate student in Computer Science and Engineering at the College of Engineering, Guindy, Chennai. His academic and professional experiences demon...

2024

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.

[1] [1]

You only look once: Unified, real -time object detection,

J. Redmon, “You only look once: Unified, real -time object detection,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2016

work page 2016

[2] [2]

A comprehensive review of yolo architectures in computer vision: From yolov1 to yolov8 and yolo -nas,

J. Terven, D. -M. Córdova-Esparza, and J. -A. Romero-González, “A comprehensive review of yolo architectures in computer vision: From yolov1 to yolov8 and yolo -nas,” Machine Learning and Knowledge Extraction, vol. 5, no. 4, pp. 1680–1716, 2023

work page 2023

[3] [3]

Complexer-yolo: Real-time 3d object detection and tracking on semantic point clouds ,

M. Simon, K. Amende, A. Kraus, J. Honer, T. Samann, H. Kaulbersch, S. Milz, and H. Michael Gross, “Complexer-yolo: Real-time 3d object detection and tracking on semantic point clouds ,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2019, pp. 0–0

work page 2019

[4] [4]

Object detection and tracking with yolo and the sliding innovation filter,

A. Moksyakov, Y. Wu, S. A. Gadsden, J. Yawney, and M. AlShabi, “Object detection and tracking with yolo and the sliding innovation filter,” Sensors, vol. 24, no. 7, p. 2107, 2024

work page 2024

[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, 2016

work page 2016

[6] [6]

Improvement of object detection based on faster r -cnn and yolo,

J. Fan, J. Lee, I. Jung, and Y. Lee, “Improvement of object detection based on faster r -cnn and yolo,” in 2021 36th International Technical Conference on Circuits/Systems, Computers and Communications (ITC-CSCC). IEEE, 2021, pp. 1–4

work page 2021

[7] [7]

SSD: Single shot multibox detector,

W. Liu, D. A nguelov, D. Erhan, C. Szegedy, S. Reed, C. -Y. Fu, and A. C. Berg, “SSD: Single shot multibox detector,” in Computer Vision –ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11–14, 2016, Proceedings, Part I 14. Springer, 2016, pp. 21–37

work page 2016

[8] [8]

Focal loss for dense object detection,

T.-Y. Ross and G. Dollár, “Focal loss for dense object detection,” in proceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 2980–2988

work page 2017