Pith. sign in

REVIEW 3 major objections 4 minor 45 references

HieraEdgeNet: A Multi-Scale Edge-Enhanced Framework for Automated Pollen Recognition

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read HieraEdgeNet claims that explicitly fusing multi-scale edge features with semantic features is what lets a compact CNN detector beat larger Transformer detectors on small, blurred pollen grains.

desk verdict A competent applied detector with a plausible but unverified headline; the 46-vs-120 inconsistency and synthetic validation need fixing, but the paper deserves review. read the letter →

arxiv 2506.07637 v1 pith:NRVNUY57 submitted 2025-06-09 cs.CV cs.LG

Yuchong Long , Wen Sun , Ningxiao Sun , Wenxiao Wang , Chao Li , Shan Yin This is my paper
classification cs.CVcs.LG
keywords HieraEdgeNetpollenrecognitionobjectdetectionenhancementmulti-scalefeaturefusionomni-kernelsmallsyntheticmicroscopydataset
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

HieraEdgeNet is a CNN object detector built for pollen grains, which are small, have soft boundaries, and sit on cluttered microscope backgrounds. The paper's claim is that explicit edge information, extracted at several scales and fused with semantic features, is what closes the accuracy gap on these targets. The evidence is a 120-class pollen detection benchmark on which HieraEdgeNet reaches mAP@.5:.95 of 0.8444, ahead of YOLOv12n (0.8315), YOLOv11n (0.8235), and RT-DETR-R18 (0.8074). The same backbone with an RT-DETR decoder reaches 0.8492, and a pruned variant keeps 0.8363 at 403 FPS. That would matter because automated pollen counting underpins paleoclimate, biodiversity, and allergy monitoring, where precision is more valuable than raw speed.

What carries the argument

The load-bearing mechanism is a three-part edge-processing pipeline. The Hierarchical Edge Module applies a fixed Sobel convolution to an early feature map and max-pools it into a three-level edge pyramid at 1/4, 1/8, and 1/16 resolution, giving each backbone stage an explicit boundary prior. The Synergistic Edge Fusion module concatenates each edge map with the matching semantic feature map and fuses them through 1x1, 3x3, and 1x1 convolutions, so boundary cues are re-injected at every scale. The Cross Stage Partial Omni-Kernel Module splits channels along the CSP idea, runs one branch through an Omni-Kernel operator made of parallel anisotropic large-kernel depthwise convolutions plus spatial, frequency-channel, and frequency-gating attention, and rejoins the skip branch; it is placed where P2 detail, P3 fused features, and upsampled P5 context meet, which is where small-object detail is richest.

What would settle it

Evaluate HieraEdgeNet and YOLOv12n on a held-out set of genuine, palynologist-annotated bright-field pollen slides; if HieraEdgeNet's mAP@.5:.95 does not remain above YOLOv12n's on real captures, the reported edge-enhancement gains are artifacts of synthetic validation.

Watch

Extended reading notes

Core claim

The central discovery, on the paper's own terms, is that a detector does not have to choose between semantic richness and boundary precision: it can have both by making edges a first-class, scale-structured feature. HieraEdgeNet's Hierarchical Edge Module turns Sobel gradients of early features into an edge pyramid, its Synergistic Edge Fusion modules inject those edges at the P3, P4, and P5 semantic levels, and its Cross Stage Partial Omni-Kernel Module refines the detail-rich P3 level with anisotropic large kernels and mixed-domain attention. The paper reports that this combination outperforms similarly sized real-time detectors and even pairs well with a Transformer decoder, which it reads as evidence that the backbone learns unusually transferable features. The quantitative core is mAP@.5:.95 = 0.8444 on 120 pollen classes, versus 0.8315 for YOLOv12n, 0.8235 for YOLOv11n, and 0.8074 for RT-DETR-R18, with mAP@.5 = 0.9501.

Load-bearing premise

Validation images are produced by the same synthetic pipeline that builds training images, so the reported accuracy assumes those composites faithfully represent real pollen microscopy.

Editorial extensions

If this is right

  • If the reported numbers hold, edge-aware multi-scale fusion is a cheap, CNN-compatible way to close the gap with Transformer detectors on small, blurred objects.
  • The backbone's transfer to an RT-DETR decoder suggests the learned features are reusable across detection heads, not tuned to one architecture.
  • A pruned version retaining 0.8363 mAP at 403 FPS implies the edge-enhancement gains survive compression, making real-time pollen monitoring plausible.
  • Sharper edge-focused activation maps, if they generalize, should translate into better localization on other microscopy targets with indistinct boundaries.

Reading between the lines

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

  • The same HEM and SEF recipe could be dropped into other small-object detectors that already use FPN-style necks, which would test whether the gain is additive and transferable.
  • A natural stress test is to retrain on only the real annotated pollen images and compare against synthetic-augmented training; the ablation would separate edge-enhancement gains from synthetic-data gains.
  • Because the paper names Z-stack data as a limitation, extending the edge pyramid to 3D volumes is an obvious next step, and the architecture seems designed for such extension.
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 / 4 minor

Summary. The paper proposes HieraEdgeNet, a CNN object-detection architecture for pollen recognition built from three modules: a Hierarchical Edge Module (HEM) that extracts multi-scale Sobel-based edge maps, Synergistic Edge Fusion (SEF) modules that combine these edge maps with backbone semantic features, and a Cross Stage Partial Omni-Kernel Module (CSPOKM) applied at the P3 feature level. On a newly constructed 120-class pollen detection dataset, the authors report mAP@.5:.95 = 0.8444 for HieraEdgeNet, outperforming YOLOv11n, YOLOv12n, RT-DETR-R18, and RT-DETR-R50; a variant with an RT-DETR decoder reaches 0.8492, and a pruned variant reaches 0.8363. The paper also presents Grad-CAM visualizations suggesting sharper edge-focused attention than baselines.

Significance. If confirmed, the architectural combination of explicit edge pyramids with efficient CSP/Omni-Kernel processing would be a practically useful contribution to small-object detection, and the release of the dataset and models supports reproducibility. The paper correctly credits CSPOKM to prior work [21] and explicitly acknowledges limitations such as the 2D design and potential domain shift. However, the current evidence does not establish the headline claim as stated: the 120-class evaluation is internally inconsistent with the 46-class confusion matrix presented in Figure 4, and the validation set is produced by the same synthesis pipeline as the training data. The practical significance for real pollen monitoring therefore remains to be demonstrated.

major comments (3)
  1. [Section 4.2, Figure 4, Table 2] The class count and metric values are internally inconsistent in a way that affects the central claim. Section 4.2 states that the confusion matrix in Figure 4a shows performance across 46 pollen classes, while the abstract and Section 4.1 describe a 120-class dataset and Table 2 reports mAP@.5:.95 = 0.8444. Additionally, the P-R curve in Figure 4b reports an mAP of 0.976 over all classes, which matches neither the mAP@.5 (0.9501) nor the mAP@.5:.95 (0.8444) in Table 2. The authors must state exactly which classes and which split were used for each reported number, and reconcile the 46-class figure with the 120-class dataset. The headline result is not verifiable until this is resolved.
  2. [Section 4.1] The validation set is generated by the same synthetic pipeline used for training: single-grain classification images are programmatically embedded onto authentic backgrounds. Consequently, the validation distribution is not independent of the training distribution, and the measured mAP may overestimate performance on real microscopy slides where natural aggregation, overlapping grains, and multi-focal-plane effects occur. The paper's own conclusion acknowledges this domain-shift limitation, but the central claim of high-accuracy automated pollen recognition on a large-scale dataset needs at least one external validation set with real annotated slides, or a clearly justified argument for treating the synthetic distribution as the target domain.
  3. [Table 2] No error bars or repeated runs are reported for any model. The headline improvements over YOLOv12n and YOLOv11n are 1.29 and 2.09 mAP@.5:.95 percentage points, respectively; without variance estimates or multiple seeds, the reader cannot assess whether these differences are statistically meaningful. At minimum, the authors should report mean and standard deviation over three or more runs, or otherwise justify why single-run results are sufficient for the claimed superiority.
minor comments (4)
  1. [Equations (2) and (3)] The concatenation operator is introduced in the text as the symbol L, but the equations use a different symbol that appears as an encircled operator; please use one consistent notation for channel-wise concatenation throughout the paper.
  2. [Table 2] The inference speed (FPS) is reported without specifying the GPU model, batch size, software framework, or whether TensorRT / half-precision inference is used. Please provide this information to make the efficiency comparisons reproducible.
  3. [Reference [44]] The LAMP pruning method is cited to a paper about an FPGA accelerator (LAMPS); this appears to be the wrong reference. Please cite the original layer-wise adaptive magnitude pruning work or provide a correct reference for the pruning technique actually used.
  4. [Section 3.4] The phrase 'area=1 signifies global regional attention' is confusing; an area parameter of 1 would usually mean a single spatial location, not a global attention region. Please clarify how the area parameter is interpreted in the AAttn module.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the mAP results are direct measurements on a held-out validation split, the CSPOKM module is explicitly credited to prior external work, and no fitted parameter is renamed as a prediction.

full rationale

The paper's central claims are empirical benchmarks, not derivations from fitted inputs. HieraEdgeNet's mAP values in Table 2 are obtained by training the model on an 80% training split and evaluating on a separate 20% validation split (Section 4.1), which is a standard measured evaluation rather than a prediction forced by construction. The three proposed modules (HEM, SEF, CSPOKM) are described as architectural components; CSPOKM is explicitly attributed to prior work [21], and the paper does not derive its performance from the definition of the module. No self-citation chain is load-bearing, as none of the cited module sources overlaps with the present authors. The main legitimate concerns raised by the skeptical review are dataset-validity issues: the validation images are generated by the same synthesis pipeline as training, and the paper contains an unresolved discrepancy between the claimed 120 classes and the 46 classes shown in the confusion matrix. These concerns affect whether the measured accuracy generalizes to real microscopy slides, but they do not make the reported result circular, because the accuracy is not defined in terms of the training loss or fitted parameters. The paper explicitly acknowledges that performance may be challenged by domain shifts between training data and real-world samples, reinforcing that no overclaim by construction is embedded in the method. Therefore, no specific circular step can be exhibited, and the correct circularity score is 0.

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

The central result is empirical, so the ledger records the explicit modeling choices (split ratio, loss hyperparameters) and the unproven data assumptions (synthetic embeddings as a proxy for real microscopy, label consistency across merged datasets, Sobel features as a valid edge prior). No new physical or procedural entities are introduced beyond the named network modules.

free parameters (4)
  • CSPOKM split ratio e = 0.25
    Section 3.3 sets e=0.25 empirically to partition channels into processed and skip paths; the architecture depends on this choice.
  • Focal loss parameters (alpha, gamma) = not reported
    Section 3.5 Eq. (4) uses alpha and gamma but the paper never states their values; chosen by hand.
  • Loss weights (lambda_reg, lambda_cls) = not reported
    Section 3.5 Eq. (5) weights regression and classification losses; values are not given.
  • AAttn area parameter = 1
    Section 3.4 states area=1 signifies global regional attention in ABlock; a design choice.
assumptions (3)
  • domain assumption The synthetic embedding pipeline generates detection images whose distribution matches real microscopy well enough for training and evaluation.
    Section 4.1 constructs the 120-class dataset by programmatically pasting single-grain images onto authentic backgrounds; the paper does not report a real-microscopy test set or domain-shift analysis.
  • domain assumption Merging several public datasets yields a consistent 120-class taxonomy with correct, compatible annotations.
    Section 2.3 and Section 4.1 aggregate multiple public datasets; inter-dataset label consistency and annotation quality are not audited.
  • domain assumption The fixed Sobel operator on feature maps is a meaningful edge prior for pollen boundaries at the resolutions used.
    Section 3.1 assumes gradient-based edges from SobelConv at early backbone stages improve localization; no quantitative edge-quality validation is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HieraEdgeNet: A Multi-Scale Edge-Enhanced Framework for Automated Pollen Recognition." pith.science (2026). https://pith.science/paper/NRVNUY57

@misc{pith2026250607637,
  author       = {Pith},
  title        = {Pith review of: HieraEdgeNet: A Multi-Scale Edge-Enhanced Framework for Automated Pollen Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NRVNUY57}},
  note         = {Machine review of arXiv:2506.07637}
}
read the original abstract

Automated pollen recognition is vital to paleoclimatology, biodiversity monitoring, and public health, yet conventional methods are hampered by inefficiency and subjectivity. Existing deep learning models often struggle to achieve the requisite localization accuracy for microscopic targets like pollen, which are characterized by their minute size, indistinct edges, and complex backgrounds. To overcome this limitation, we introduce HieraEdgeNet, a multi-scale edge-enhancement framework. The framework's core innovation is the introduction of three synergistic modules: the Hierarchical Edge Module (HEM), which explicitly extracts a multi-scale pyramid of edge features that corresponds to the semantic hierarchy at early network stages; the Synergistic Edge Fusion (SEF) module, for deeply fusing these edge priors with semantic information at each respective scale; and the Cross Stage Partial Omni-Kernel Module (CSPOKM), which maximally refines the most detail-rich feature layers using an Omni-Kernel operator - comprising anisotropic large-kernel convolutions and mixed-domain attention - all within a computationally efficient Cross-Stage Partial (CSP) framework. On a large-scale dataset comprising 120 pollen classes, HieraEdgeNet achieves a mean Average Precision (mAP@.5) of 0.9501, significantly outperforming state-of-the-art baseline models such as YOLOv12n and RT-DETR. Furthermore, qualitative analysis confirms that our approach generates feature representations that are more precisely focused on object boundaries. By systematically integrating edge information, HieraEdgeNet provides a robust and powerful solution for high-precision, high-efficiency automated detection of microscopic objects.

Figures

Figures reproduced from arXiv: 2506.07637 by the authors.

Figure 1
Figure 1. The three core modules of HieraEdgeNet: (a) The HEM explicitly extracts and organizes multi-scale edge [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. The complete architecture of HieraEdgeNet, featuring enhancements for small objects, complex background [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. The distribution of different pollen types in the dataset, highlighting the diversity and scale of the training [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Detailed performance validation of the proposed HieraEdgeNet model. (a) The confusion matrix for all pollen [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: Visual analysis of HieraEdgeNet’s enhanced edge perception. Grad-CAM heatmaps compare the activation [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

45 extracted references · 41 canonical work pages

  1. [21]

    YOLO-CSPOKM: Efficient small object detector via CSP omni-kernel model

    Wu Wei. YOLO-CSPOKM: Efficient small object detector via CSP omni-kernel model. In Proceedings of the 4th Asia-pacific Artificial Intelligence and Big Data Forum, pages 716–721. Association for Computing Machinery, New York, NY , USA, April 2025

    work page 2025

  2. [1]

    Pollen and spores as proxies for palaeoenvironment reconstruction: A review of sediment-based research

    Mohammad Abass Zargar, Muzafar Riyaz, Shaik Mohammad Hussain, Arfat Nazir, Saima Hamid, Ishfaq Ahmad Gojree, and Khursheed Ahmad Parray. Pollen and spores as proxies for palaeoenvironment reconstruction: A review of sediment-based research. Journal of the Palaeontological Society of India, page 5529360251318072, April 2025

    work page 2025

  3. [2]

    Impact of biodiversity loss on pollen allergies: A bibliometric analysis

    Karoline Braun and Daniela Haluza. Impact of biodiversity loss on pollen allergies: A bibliometric analysis. Sustainability, 16(21):9285, 2024

    work page 2024

  4. [3]

    Beggs, Darren Wraith, and Janet M

    Divya Dwarakanath, Andelija Milic, Paul J. Beggs, Darren Wraith, and Janet M. Davies. A global survey addressing sustainability of pollen monitoring. World Allergy Organization Journal, 17(12):100997, 2024

    work page 2024

  5. [4]

    Palyno-morphological study of allergenic flora of samarkand, uzbekistan

    Zamira Jumayeva and Aziza Nozimova. Palyno-morphological study of allergenic flora of samarkand, uzbekistan. American Journal of Plant Sciences, 14(5):533–541, 2023

    work page 2023

  6. [5]

    Markus Berger, Katharina Bastl, Maximilian Bastl, Lukas Dirr, and Uwe E. Berger. Digital Health for Patients with Pollen Allergy: The Pollen App Experience and Beyond, pages 49–61. Springer Nature Switzerland, Cham, 2025

    work page 2025

  7. [6]

    The influence of environmental pollution on the allergenic potential of grass pollen

    Ivana Prodi´c, Rajna Mini´c, and Marija Stojadinovi´c. The influence of environmental pollution on the allergenic potential of grass pollen. Aerobiologia, 41(1):3–16, March 2025

    work page 2025

  8. [7]

    García-Orellana, Antonio García-Manso, Horacio M

    Ramón Gallardo-Caballero, Carlos J. García-Orellana, Antonio García-Manso, Horacio M. González-Velasco, Rafael Tormo-Molina, and Miguel Macías-Macías. Precise pollen grain detection in bright field microscopy using deep learning techniques. Sensors, 19(16):3583, 2019

    work page 2019

Show all 45 references
  1. [8]

    Pollen grain classification using some convolutional neural network architectures

    Benjamin Garga, Hamadjam Abboubakar, Rodrigue Saoungoumi Sourpele, David Libouga Li Gwet, and Laurent Bitjoka. Pollen grain classification using some convolutional neural network architectures. Journal of Imaging, 10(7):158, July 2024

    2024

  2. [9]

    Deep learning

    Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. Nature, 521(7553):436–444, 2015

    2015

  3. [10]

    F. M. Shamrat, Mohd Yamani Idna Idris, Xujuan Zhou, Majdi Khalid, Sharmin Sharmin, Zeseya Sharmin, Kawsar Ahmed, and Mohammad Ali Moni. PollenNet: A novel deep learning architecture for high precision pollen grain classification through deep learning and explainable AI. Heliyo...

    2024

  4. [11]

    Landsmeer, Emile A

    Sander H. Landsmeer, Emile A. Hendriks, Letty A. De Weger, Johan H.C. Reiber, and Berend C. Stoel. Detection of pollen grains in multifocal optical microscopy images of air samples. Microscopy Research and Technique, 72(6):424–430, June 2009

    2009

  5. [12]

    Nhat Rich Nguyen, Matina Donalson-Matasci, and Min C. Shin. Improving pollen classification with less training effort. In 2013 IEEE Workshop on Applications of Computer Vision (WACV), pages 421–426, January 2013

    2013

  6. [13]

    Víctor Marcos, Gabriel Cristóbal, Tomás Rodríguez, Amelia Gonzalez-Porto, Cristina Pardo, and Oscar Déniz

    Rafael Redondo, Gloria Bueno, François Chung, Rodrigo Nava, J. Víctor Marcos, Gabriel Cristóbal, Tomás Rodríguez, Amelia Gonzalez-Porto, Cristina Pardo, and Oscar Déniz. Pollen segmentation and feature evaluation for automatic classification in bright-field microscopy. Compute...

    2015

  7. [14]

    YOLOv12: Attention-centric real-time object detectors, February 2025

    Yunjie Tian, Qixiang Ye, and David Doermann. YOLOv12: Attention-centric real-time object detectors, February 2025

    2025

  8. [15]

    Feature pyramid networks for object detection

    Tsung-Yi Lin, Piotr Dollár, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2117–2125, 2017

    2017

  9. [16]

    Detrs beat yolos on real-time object detection

    Yian Zhao, Wenyu Lv, Shangliang Xu, Jinman Wei, Guanzhong Wang, Qingqing Dang, Yi Liu, and Jie Chen. Detrs beat yolos on real-time object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16965–16974, 2024

    2024

  10. [17]

    End-to-end object detection with transformers

    Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision – ECCV 2020, volume 12346...

    2020

  11. [18]

    Detection and recognition of pollen grains in multilabel microscopic images

    El˙zbieta Kubera, Agnieszka Kubik-Komar, Paweł Kurasi´nski, Krystyna Piotrowska-Weryszko, and Magdalena Skrzypiec. Detection and recognition of pollen grains in multilabel microscopic images. Sensors, 22(7):2690, January 2022

    2022

  12. [19]

    PollenDetect: An open-source pollen viability status recognition system based on deep learning neural networks

    Zhihao Tan, Jing Yang, Qingyuan Li, Fengxiang Su, Tianxu Yang, Weiran Wang, Alifu Aierxi, Xianlong Zhang, Wanneng Yang, Jie Kong, and Ling Min. PollenDetect: An open-source pollen viability status recognition system based on deep learning neural networks. International journal...

    2022

  13. [20]

    Holistically-nested edge detection

    Saining Xie and Zhuowen Tu. Holistically-nested edge detection. https://arxiv.org/abs/1504.06375v2, April 2015

    2015 arXiv

  14. [22]

    CSPNet: A new backbone that can enhance learning capability of CNN, November 2019

    Chien-Yao Wang, Hong-Yuan Mark Liao, I.-Hau Yeh, Yueh-Hua Wu, Ping-Yang Chen, and Jun-Wei Hsieh. CSPNet: A new backbone that can enhance learning capability of CNN, November 2019

    2019

  15. [23]

    Batch normalization: Accelerating deep network training by reducing internal covariate shift

    Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International Conference on Machine Learning, pages 448–456. pmlr, 2015

    2015

  16. [24]

    Venkata Rami Reddy

    Mupparaju Sohan, Thotakura Sai Ram, and Ch. Venkata Rami Reddy. A review on YOLOv8 and its advancements. In I. Jeena Jacob, Selwyn Piramuthu, and Przemyslaw Falkowski-Gilski, editors, Data Intelligence and Cognitive Informatics, pages 529–545. Springer Nature Singapore, Singap...

    2024

  17. [25]

    Weinberger

    Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q. Weinberger. Densely connected convolutional networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4700–4708, 2017

    2017

  18. [26]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, and Sylvain Gelly. An image is worth 16x16 words: Transformers for image recognition at scale. Arxiv Preprint Arxiv:20...

    2010 arXiv

  19. [27]

    Ultralytics/yolov5: V3

    Glenn Jocher, Alex Stoken, Jirka Borovec, Liu Changyu, Adam Hogan, Laurentiu Diaconu, Jake Poznanski, Lijun Yu, Prashant Rai, and Russ Ferriday. Ultralytics/yolov5: V3. 0. Zenodo, 2020

    2020

  20. [28]

    A comparative study of YOLOv5 and YOLOv7 object detection algorithms

    Oluwaseyi Ezekiel Olorunshola, Martins Ekata Irhebhude, and Abraham Eseoghene Evwiekpaefe. A comparative study of YOLOv5 and YOLOv7 object detection algorithms. Journal of Computing and Social Informatics , 2(1):1–12, 2023

    2023

  21. [29]

    A 3x3 isotropic gradient operator for image processing

    Irwin Sobel and Gary Feldman. A 3x3 isotropic gradient operator for image processing. A Talk at the Stanford Artificial Project in, 1968:271–272, 1968

    1968

  22. [30]

    No more strided convolutions or pooling: A new CNN building block for low-resolution images and small objects, August 2022

    Raja Sunkara and Tie Luo. No more strided convolutions or pooling: A new CNN building block for low-resolution images and small objects, August 2022. 15

    2022

  23. [31]

    Omni-kernel network for image restoration

    Yuning Cui, Wenqi Ren, and Alois Knoll. Omni-kernel network for image restoration. Proceedings of the AAAI Conference on Artificial Intelligence, 38(2):1426–1434, March 2024

    2024

  24. [32]

    Schneider, and Steffen Viken Valvag

    Åge Kvalnes, Dag Johansen, Robbert van Renesse, Fred B. Schneider, and Steffen Viken Valvag. Omni-kernel: An operating system architecture for pervasive monitoring and scheduling. IEEE Transactions on Parallel and Distributed Systems, 26(10):2849–2862, 2014

    2014

  25. [33]

    Scaling up your kernels to 31x31: Revisiting large kernel design in cnns

    Xiaohan Ding, Xiangyu Zhang, Jungong Han, and Guiguang Ding. Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11963–11975, 2022

    2022

  26. [34]

    Fcanet: Frequency channel attention networks

    Zequn Qin, Pengyi Zhang, Fei Wu, and Xi Li. Fcanet: Frequency channel attention networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 783–792, 2021

    2021

  27. [35]

    Squeeze-and-excitation networks

    Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7132–7141, 2018

    2018

  28. [36]

    Global filter networks for image classification

    Yongming Rao, Wenliang Zhao, Zheng Zhu, Jiwen Lu, and Jie Zhou. Global filter networks for image classification. Advances in Neural Information Processing Systems, 34:980–993, 2021

    2021

  29. [37]

    POLLEN73S: An image dataset for pollen grains classification

    Gilberto Astolfi, Ariadne Barbosa Goncalves, Geazy Vilharva Menezes, Felipe Silveira Brito Borges, Angelica Christina Melo Nunes Astolfi, Edson Takashi Matsubara, Marco Alvarez, and Hemerson Pistori. POLLEN73S: An image dataset for pollen grains classification. Ecological Info...

    2020

  30. [38]

    Combating data incompetence in pollen images detection and classification for pollinosis prevention

    Natalia Khanzhina, Andrey Filchenkov, Natalia Minaeva, Larisa Novoselova, Maxim Petukhov, Irina Kharisova, Julia Pinaeva, Georgiy Zamorin, Evgeny Putin, and Elena Zamyatina. Combating data incompetence in pollen images detection and classification for pollinosis prevention. Co...

    2022

  31. [39]

    Galan, J

    C. Galan, J. Oteros, A. Damialis, and F. Kolek. Cretan Pollen Dataset v1 (CPD-1). https://data.niaid.nih. gov/resources?id=zenodo_4756360, 2021

    2021

  32. [40]

    Monospecific Mediterranean Pollen Images Dataset, June 2023

    Thomas Gregor, Juraj Paule, and Patrik Zuber. Monospecific Mediterranean Pollen Images Dataset, June 2023

    2023

  33. [41]

    Automated pollen detection with an affordable technology

    Nam Cao, Matthias Meyer, Lothar Thiele, and Olga Saukh. Automated pollen detection with an affordable technology. In Ewsn, pages 108–119, 2020

    2020

  34. [42]

    YOLOv11: An overview of the key architectural enhancements, October 2024

    Rahima Khanam and Muhammad Hussain. YOLOv11: An overview of the key architectural enhancements, October 2024

    2024

  35. [43]

    Detrs with hybrid matching

    Ding Jia, Yuhui Yuan, Haodi He, Xiaopei Wu, Haojun Yu, Weihong Lin, Lei Sun, Chao Zhang, and Han Hu. Detrs with hybrid matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19702–19712, 2023

    2023

  36. [44]

    LAMPS: A layer-wised mixed-precision-and-sparsity accelerator for NAS-optimized CNNs on FPGA

    Shuxin Yang, Chenchen Ding, Mingqiang Huang, Kai Li, Chenghao Li, Zikun Wei, Sixiao Huang, Jingyao Dong, Liuyang Zhang, and Hao Yu. LAMPS: A layer-wised mixed-precision-and-sparsity accelerator for NAS-optimized CNNs on FPGA. In 2024 IEEE 32nd Annual International Symposium on...

    2024

  37. [45]

    Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra

    Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-CAM: Visual explanations from deep networks via gradient-based localization. International Journal of Computer Vision, 128(2):336–359, February 2020. 16

    2020

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.