CBAM-STN-TPS-YOLO: Enhancing Agricultural Object Detection through Spatially Adaptive Attention Mechanisms

Satvik Praveen; Yoonsung Jung

REVIEW 4 major objections 6 minor 31 references

CBAM-STN-TPS-YOLO: Enhancing Agricultural Object Detection through Spatially Adaptive Attention Mechanisms

T0 review · 4 major / 6 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 Replacing affine warping with thin-plate splines and adding CBAM attention improves YOLO's precision, recall, and mAP on agricultural datasets, cutting false positives by 12 percent.

desk verdict Sensible architectural combo, but the key PGP comparison is contaminated by a likely baseline mix-up, so the headline gain isn't reliable. read the letter →

arxiv 2506.07357 v1 pith:3EVHUF6N submitted 2025-06-09 cs.CV cs.LG

Satvik Praveen , Yoonsung Jung This is my paper

classification cs.CVcs.LG

keywords objectdetectionprecisionagriculturespatialtransformernetworksthin-platesplinesCBAMattentionocclusionhandlingYOLOedgedeployment

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 the two standard fixes for spatially rigid object detectors—affine warping and channel/spatial attention—can be pushed further in one architecture. Replacing the affine transformation inside a Spatial Transformer Network with a Thin-Plate Spline (TPS) deformation field, and then passing features through the Convolutional Block Attention Module (CBAM), improves YOLO's precision, recall, and mean average precision on occlusion-heavy agricultural datasets. The headline result is a 12% reduction in false positives on the Plant Growth and Phenotyping (PGP) dataset relative to STN-YOLO, with no loss of real-time speed. A sympathetic reader should care because the claim is concrete and testable: if true, a lightweight recipe exists for making detectors resilient to bent leaves, overlaps, and cluttered backgrounds on edge devices.

What carries the argument

The mechanism is a Thin-Plate Spline (TPS) deformation field inserted into the Spatial Transformer Network's grid generator, replacing the affine sampler, followed by CBAM's sequential channel and spatial attention. TPS maps coordinates by T(x,y)=a0+a1x+a2y+sum wi U(||(x,y)-(xi,yi)||) with kernel U(r)=$r^{2}$ log r, and is trained by minimizing bending energy with a regularization parameter λ that trades smoothness against flexibility. CBAM computes a channel mask from global average/max pooling through two fully connected layers, then a spatial mask from a 7x7 convolution over pooled features; both masks are multiplied into the feature map. The machinery's job is to let the detector align irregular plant structures before the YOLO backbone sees them and to suppress the background clutter that generates false positives.

What would settle it

Benchmark the proposed model against STN-YOLO using the original codebase with identical preprocessing, annotations, data splits, and training hyperparameters, changing only the transformation module (TPS vs affine) and attention module (CBAM vs none); if the precision and mAP margins shrink to within one standard deviation, the central claim fails. The specific place to look is TPS control-point placement and the value of lambda, since the paper leaves them unspecified.

Watch

Extended reading notes

Core claim

The central claim is that non-rigid spatial alignment plus attention-guided feature refinement is sufficient to beat both plain YOLO and affine STN-YOLO on agricultural object detection. The paper argues that affine transformations can correct global rotation, translation, and scale but cannot represent the local bending and curling of leaves; TPS adds a smooth deformation field parameterized by control points, and CBAM then reweights feature channels and spatial locations to suppress background. Across the PGP and MelonFlower datasets, CBAM-STN-TPS-YOLO reports the highest precision, recall, and mAP in every training configuration, and its inference time (about 14 ms per frame, or 46 ms on a Jetson Xavier with TensorRT) stays below the affine baseline. The paper treats the per-module ablations—STN-TPS-YOLO and CBAM-STN-YOLO—as evidence that each ingredient contributes, and a paired t-test over 100 epochs reports p<0.05 for all four metrics against STN-YOLO.

Load-bearing premise

The paper's gains depend on the assumption that replacing affine with TPS and adding CBAM are the only material changes from the STN-YOLO baseline, and that unstated implementation details—where TPS is inserted, how control points are placed, the λ value, and the PGP train/test split—do not drive the reported improvements.

Editorial extensions

If this is right

On the PGP and MelonFlower benchmarks, the full model achieves the best precision, recall, and mAP among YOLO, STN-YOLO, and the two single-module ablations.
The 12% false-positive reduction on PGP is a direct consequence of the improved precision (96.27 vs 95.34 for STN-YOLO).
The TPS regularization parameter λ offers a tunable trade-off between transformation smoothness and detection performance.
Inference remains real-time: about 14 ms per frame on an A100 GPU and about 46 ms per frame on a Jetson Xavier with TensorRT.
In six of eight augmentation scenarios, the full model has the best precision, indicating robustness to rotation, shear, and cropping distortions.

Reading between the lines

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

The same TPS-plus-CBAM recipe likely transfers to other non-rigid detection tasks (medical imaging, wildlife, packaged goods), since the targeted failure mode—rigid alignment plus background clutter—is not agriculture-specific; the paper only evaluates agriculture.
The paired t-test reported in Section IV.D compares metrics across 100 training epochs within one run, not across the three random seeds used elsewhere; a test across seeds would be a stronger check that the gains are architectural, not accidental.
Because the paper leaves TPS control-point placement and the λ value unspecified, a sensitivity analysis varying these would determine whether the benefit is from the TPS concept or from particular settings.
On MelonFlower, where the mAP is low for all models (44.9–47.2), the large precision gain suggests the modules help most on small, low-contrast objects; a size-stratified false-positive analysis would test that directly.

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

First, the useful news: replacing the affine sampler in STN-YOLO with a TPS warp and adding CBAM is a sensible thing to try, and the paper is upfront that it is extending Zambre et al. The MelonFlower results, where the proposed model beats both baselines in precision and mAP, are internally consistent, and the Jetson inference time (46 ms) is a nice practical data point.

Now the problem. The paper's central claim, that the full model substantially improves over STN-YOLO on PGP, rests on Table V. But the STN-YOLO row in Table V is numerically identical to the '28x28' row of Table II, which the caption attributes to Zambre et al., not to the authors' own runs. Meanwhile, the authors' own controlled runs appear in Table IV, where the no-augmentation STN-YOLO has 85.04 accuracy and 72.36 mAP. Those are different numbers. If Table IV is the true baseline, the proposed model's mAP gain is 0.25, not 1.15, and the two intermediate models (STN-TPS-YOLO and CBAM-STN-YOLO) actually finish below STN-YOLO in mAP. That directly contradicts the attribution of the gain to TPS and CBAM. The paired t-test in Section IV.D compares metrics across 100 training epochs; that is not a valid test of run-to-run variability and cannot rescue the comparison.

Other soft spots: no code or data split is provided, so the exact TPS control-point placement, lambda, and the PGP train/test split are unverifiable. The statistical analysis needs a proper run-level test with more than three seeds. These are fixable in a revision.

Verdict: the architecture idea is worth discussing, and the MelonFlower numbers give some reason to think it works. But as written, the headline PGP result is not established. A serious referee should ask the authors to resolve the Table IV/Table V discrepancy, re-run the PGP baseline under the same protocol, and drop the epoch-based t-test. If they can do that, the paper could become a modest but honest contribution. Without that, it shouldn't be accepted.

Recommendation: send it to peer review with a request for major revision, but tell the authors the current evidence is not adequate.

Referee Report

4 major / 6 minor

Summary. This manuscript proposes CBAM-STN-TPS-YOLO, an agricultural object detector that extends the STN-YOLO architecture by replacing the affine spatial transformation with a Thin-Plate Spline (TPS) deformation and adding a Convolutional Block Attention Module (CBAM). The authors report experiments on the PGP and MelonFlower datasets, including an augmentation robustness study (Table IV), benchmark comparisons (Table V), three-run means and standard deviations, inference timings, and a paired t-test over training epochs. The central claim is that the proposed model outperforms YOLO and STN-YOLO in precision, recall, and mAP, with roughly 12% fewer false positives.

Significance. If the claims held, the contribution would be a lightweight, modular modification to STN-YOLO with consistent small gains, useful for agricultural edge deployment. The paper has several strengths: it compares against an external baseline (Zambre et al.), includes separate ablations for TPS and CBAM in Table IV, reports three-run variability, provides inference-time measurements, and makes a falsifiable prediction about non-rigid deformation handling. It is not circular: the reported metrics are not constructed from the TPS hyperparameter. However, the significance is currently constrained by (i) an internal contradiction between the STN-YOLO baseline in Table V and the controlled baseline in Table IV, (ii) an invalid statistical test in Section IV.D, and (iii) the absence of implementation details and code. As presented, the claim of substantial improvement over STN-YOLO is not yet established.

major comments (4)

[Section IV.C, Tables II, IV, V] Table V's PGP STN-YOLO row (81.63±1.53 accuracy, 95.34±0.76 precision, 89.52±0.57 recall, 72.56±0.90 mAP) is numerically identical to the 28×28 row of Table II, whose caption attributes the numbers to Zambre et al. rather than to the authors' own runs. The paper's own controlled Table IV reports STN-YOLO with no augmentation as 85.04, 94.76, 89.67, 72.36 on the same PGP data. A baseline cannot take both values. If Table IV is the valid controlled comparison, the proposed model's no-augmentation PGP gains over STN-YOLO are only +0.16 precision, +0.73 recall, and +0.25 mAP, and the intermediate STN-TPS-YOLO and CBAM-STN-YOLO rows have lower mAP than STN-YOLO (72.27 and 72.26 vs. 72.36). The Section IV.C statement that the model 'substantially outperforms' STN-YOLO is therefore not supported by the controlled evidence as presented.
[Section IV.D] The paired t-test in Section IV.D is computed over the 100 training epochs for each metric. Epoch-to-epoch values within a single training run are not independent samples, and the test does not compare independent model runs; it can at most show that the training trajectories differ, not that one model generalizes better. The authors already report three-run means and standard deviations in Tables IV and V; a valid significance test would use those run-level results (or more runs) rather than per-epoch values. Without this, the sentence in Section IV.D claiming the observed metric gains are 'not incidental' is unsupported.
[Section III.G, Section III.H, Fig. 4] The method section never specifies how the TPS is integrated into the STN: where the TPS sampler replaces the affine grid generator, how the N control points in Eq. (1) are chosen, what value the regularization parameter λ in Eq. (2) takes, or how the localization network outputs the affine coefficients and TPS weights. Equation (1) is a standard TPS coordinate map, but the differentiable end-to-end training procedure is not described. Since Section III.K says the architecture is based on the STN-YOLO codebase and no code or data split is released, the reported results are not reproducible and the attribution of the gains to TPS plus CBAM cannot be independently checked. Please provide these implementation details and, ideally, release the code.
[Abstract, Section III, Section IV] The abstract promises an examination of the TPS regularization parameter λ and claims a 12% reduction in false positives, but neither is present in the experimental section: no λ sweep or result is reported, and no table or text computes false-positive counts. The Section IV.B statement about reducing false positives is based only on higher precision, which is not the same as an FP count. Either add the missing λ experiments and a false-positive analysis (for example, derived from the confusion matrix in Fig. 5) or remove these claims from the abstract.

minor comments (6)

[Section II.E, Fig. 4] The text says the STN aligns features and then the TPS transformation 'follows, replacing affine transformations,' while Fig. 4 appears to place the TPS module inside the STN module; please clarify whether TPS replaces the STN sampler or is a separate stage.
[Section III.D] The numerical example in Section III.D uses randomized TPS weights and reports an arbitrary output coordinate; it does not illustrate the training or regularization procedure and should be removed or replaced with a concrete example tied to Eq. (2).
[Section IV.A] Section IV.A lists GlobalWheat2020, PlantDoc, and MelonFlower as benchmark datasets, but Table V reports results only for PGP and MelonFlower; please either add the missing results or remove the unused datasets from the experimental setup.
[Eq. (2)] Equation (2) uses f in the bending-energy integral but T in the data term; please define the relationship between f and T and use consistent notation.
[Table V] The 'Accuracy' metric in Tables IV and V is not defined for an object detection task; please state whether it is per-image classification accuracy or another quantity, since this affects interpretation of the reported numbers.
[References] Reference [7] is a 2004 technical report on non-rigid registration rather than an authoritative source for the limitations of affine STNs, and several references (e.g., [1]) are informal preprints; please consider replacing them with peer-reviewed sources.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the empirical comparison is self-contained and not definitionally tied to its inputs.

full rationale

The paper's central claim is empirical: that adding TPS and CBAM to STN-YOLO improves precision, recall, and mAP on PGP and MelonFlower. No equation in the paper defines the reported metrics in terms of the proposed modules' parameters; the metrics come from benchmark evaluation on held-out data. The only studied parameter, the TPS regularization lambda, is a hyperparameter examined in ablations, and the final metrics are not constructed from it. The STN-YOLO baseline is external to this paper, and even the numerical coincidence between Table V's PGP STN-YOLO row and Table II's 28x28 row from Zambre et al. is a data-provenance or reproducibility concern, not a reduction of the claimed result to its inputs. No self-citation is load-bearing because the authors do not rely on their own prior theorems, and no ansatz is smuggled in via citation in a way that makes the conclusion equivalent to the premise. Therefore no circular step meets the required quote-and-reduction standard, and the appropriate finding is no significant circularity.

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

The paper introduces no new physical or mathematical entities. Its dependencies are all borrowed or underspecified: known modules (STN, TPS, CBAM, YOLO), an assembled dataset with SAM-based labels, and an unreported lambda and control-point configuration. The count of free parameters reflects the unstated implementation choices that a replication would need.

free parameters (3)

TPS regularization parameter lambda = not reported
Equation (2) defines lambda as the smoothness-regularization weight; the text says it is critical for balancing smoothness and detection performance but gives no value or sweep table, so the exact trained model is underspecified.
TPS control-point count N = not reported
Equation (1) sums over N control points that define the deformation field; the paper does not say how control points are chosen or how many are used in the YOLO feature maps.
STN localization feature map size = 28x28 (inherited from the STN-YOLO baseline)
Table II reproduces the baseline's localization output sweep and indicates 28x28 was best, but the proposed model's localization output is not stated.

assumptions (4)

domain assumption The pseudo-RGB mapping of the 580/660/730 nm bands preserves the visual information needed for object detection.
Section III.A.2 describes creating pseudo-RGB for pretrained models but gives no analysis that this mapping retains discriminative plant features.
domain assumption Labels produced by Segment Anything plus manual refinement are accurate enough to train and evaluate detectors.
Section III.A.3 uses these labels as ground truth for all models; label noise is assumed to affect models equally.
domain assumption The STN-YOLO training protocol transfers unchanged when TPS and CBAM are inserted.
Section IV.A reuses the STN-YOLO setup and hyperparameters; no per-module tuning or stability analysis is reported.
ad hoc to paper Validation metrics over the 100 training epochs are independent samples suitable for paired t-tests.
Section IV.D performs a paired t-test across epochs; this independence assumption is false for training trajectories and is not defended.

how reviews work

0 comments

Cite this review

Pith. "Pith review of CBAM-STN-TPS-YOLO: Enhancing Agricultural Object Detection through Spatially Adaptive Attention Mechanisms." pith.science (2026). https://pith.science/paper/3EVHUF6N

@misc{pith2026250607357,
  author       = {Pith},
  title        = {Pith review of: CBAM-STN-TPS-YOLO: Enhancing Agricultural Object Detection through Spatially Adaptive Attention Mechanisms},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3EVHUF6N}},
  note         = {Machine review of arXiv:2506.07357}
}

read the original abstract

Object detection is vital in precision agriculture for plant monitoring, disease detection, and yield estimation. However, models like YOLO struggle with occlusions, irregular structures, and background noise, reducing detection accuracy. While Spatial Transformer Networks (STNs) improve spatial invariance through learned transformations, affine mappings are insufficient for non-rigid deformations such as bent leaves and overlaps. We propose CBAM-STN-TPS-YOLO, a model integrating Thin-Plate Splines (TPS) into STNs for flexible, non-rigid spatial transformations that better align features. Performance is further enhanced by the Convolutional Block Attention Module (CBAM), which suppresses background noise and emphasizes relevant spatial and channel-wise features. On the occlusion-heavy Plant Growth and Phenotyping (PGP) dataset, our model outperforms STN-YOLO in precision, recall, and mAP. It achieves a 12% reduction in false positives, highlighting the benefits of improved spatial flexibility and attention-guided refinement. We also examine the impact of the TPS regularization parameter in balancing transformation smoothness and detection performance. This lightweight model improves spatial awareness and supports real-time edge deployment, making it ideal for smart farming applications requiring accurate and efficient monitoring.

Figures

Figures reproduced from arXiv: 2506.07357 by the authors.

**Figure 1.** CBAM-STN-TPS-YOLO Model Pipeline Spatial Transformer Networks (STNs) [6] enhance deep arXiv:2506.07357v1 [cs.CV] 9 Jun 2025 [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗

**Figure 3.** Visual example of Thin-Plate Spline (TPS) warping applied to a plant image. The input image (left) shows the original plant with grid lines, while the warped image (right) demonstrates the non-rigid transformation learned using TPS. This visualization illustrates the TPS module’s ability to model smooth spatial deformations, which are critical for aligning irregular plant structures. C. Procedure for CBAM-STN-TPS-YO… view at source ↗

**Figure 2.** Example images from PGP dataset. The top rows (2a [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: Architecture of the proposed CBAM-STN-TPS-YOLO model. The input image is first processed by the Spatial [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

**Figure 5.** Figure 5: Confusion matrix showing the classification perfor [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

**Figure 6.** Figure 6: Label correlation heatmap visualizing inter-class rela [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

**Figure 7.** Figure 7: Sample predictions from a validation batch on the PGP [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 24 canonical work pages

[1]

Yolo deep learning algorithm for object detection in agriculture: a review,

A. Singh and R. Sharma, “Yolo deep learning algorithm for object detection in agriculture: a review,”ResearchGate, 2024. [Online]. Available: https://www.researchgate.net/publication/387051427 YOLO deep learning algorithm for object detection in agriculture a review

work page arXiv 2024
[2]

Advancements in yolo for agricultural object detection,

J. Li and H. Wang, “Advancements in yolo for agricultural object detection,”arXiv, 2024. [Online]. Available: https://arxiv.org/abs/2401. 10379

work page 2024
[3]

Precise Apple Detection and Localization in Orchards using YOLOv5 for Robotic Harvesting Systems

J. Ziyue, Y . Bo, and L. Boyun, “Precise apple detection and localization in orchards using yolov5 for robotic harvesting systems,” 2024. [Online]. Available: https://arxiv.org/abs/2405.06260

work page Pith review arXiv 2024
[4]

Artificial intelligence in agriculture market size report, 2030,

Grand View Research, “Artificial intelligence in agriculture market size report, 2030,” 2023, accessed: 2025-04-02. [On- line]. Available: https://www.grandviewresearch.com/industry-analysis/ artificial-intelligence-in-agriculture-market

work page 2023
[5]

Spatial Transformer Network YOLO Model for Agricultural Object Detection

Y . Zambre, E. Rajkitkul, A. Mohan, and J. Peeples, “Spatial transformer network yolo model for agricultural object detection,” 2024. [Online]. Available: https://arxiv.org/abs/2407.21652

work page Pith review arXiv 2024
[6]

Spatial transformer networks,

M. Jaderberg, K. Simonyan, A. Zisserman, and K. Kavukcuoglu, “Spatial transformer networks,”Advances in Neural Information Processing Systems (NeurIPS), 2015. [Online]. Available: https: //papers.neurips.cc/paper/5854-spatial-transformer-networks.pdf

work page 2015
[7]

Limitations of affine transformations in spatial transformer networks,

R. Brown and Y . Sun, “Limitations of affine transformations in spatial transformer networks,”Princeton Graphics Group,

work page
[8]

Smooth extrapolation of unknown anatomy via statistical shape models,

R. B. Grupp, H. Chiang, Y . Otake, R. J. Murphy, C. R. Gordon, M. Armand, and R. H. Taylor, “Smooth extrapolation of unknown anatomy via statistical shape models,” inMedical Imaging 2015: Image-Guided Procedures, Robotic Interventions, and Modeling, R. J. Webster and Z. R. Yaniv, Eds., vol. 9415. SPIE, Mar. 2015, p. 941524. [Online]. Available: http://dx....

work page doi:10.1117/12.2081310 2015

Show all 31 references

[9]

Cbam: Convolutional block attention module,

S. Woo, J. Park, J.-Y . Lee, and I. S. Kweon, “Cbam: Convolutional block attention module,” 2018. [Online]. Available: https://arxiv.org/ abs/1807.06521

2018 arXiv
[10]

Comparison of single-shot and two- shot deep neural network models for whitefly detection in iot web application,

C. U. Parab, C. Mwitta, M. Hayes, J. M. Schmidt, D. Riley, K. Fue, S. Bhandarkar, and G. C. Rains, “Comparison of single-shot and two- shot deep neural network models for whitefly detection in iot web application,”AgriEngineering, vol. 4, pp. 507–522, 2022

2022
[11]

Deep learning for generic object detection: A survey,

L. Liu, W. Ouyang, X. Wang, P. Fieguth, J. Chen, X. Liu, and M. Pietik ¨ainen, “Deep learning for generic object detection: A survey,” International journal of computer vision, vol. 128, pp. 261–318, 2020

2020
[12]

Recent advances in deep learning for object detection,

X. Wu, D. Sahoo, and S. C. Hoi, “Recent advances in deep learning for object detection,”Neurocomputing, vol. 396, pp. 39–64, 2020

2020
[13]

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,” inAdvances in Neural Information Processing Systems, C. Cortes, N. Lawrence, D. Lee, M. Sugiyama, and R. Garnett, Eds., vol. 28. Curran Associates, Inc., 2015

2015
[14]

Ty-net: Transforming yolo for hand gesture recognition,

Z. Do ˇzdor, Z. Kalafati ´c, ˇZ. Ban, and T. Hrka ´c, “Ty-net: Transforming yolo for hand gesture recognition,”IEEE access, 2023

2023
[15]

Yolov3: An incremental improvement,

J. Redmon and A. Farhadi, “Yolov3: An incremental improvement,” 2018

2018
[16]

A locally- constrained yolo framework for detecting small and densely-distributed building footprints,

Y . Xie, J. Cai, R. Bhojwani, S. Shekhar, and J. Knight, “A locally- constrained yolo framework for detecting small and densely-distributed building footprints,”International Journal of Geographical Information Science, vol. 34, no. 4, pp. 777–801, 2020

2020
[17]

Crop disease detection using yolo,

A. Morbekar, A. Parihar, and R. Jadhav, “Crop disease detection using yolo,” in2020 International Conference for Emerging Technology (INCET), 2020, pp. 1–5

2020
[18]

A yolo-based pest detection system for precision agriculture,

M. Lippi, N. Bonucci, R. F. Carpio, M. Contarini, S. Speranza, and A. Gasparri, “A yolo-based pest detection system for precision agriculture,” in2021 29th Mediterranean Conference on Control and Automation (MED), 2021, pp. 342–347

2021
[19]

Real- time object detection and robotic manipulation for agriculture using a yolo-based learning approach,

H. Zhao, Z. Tang, Z. Li, Y . Dong, Y . Si, M. Lu, and G. Panoutsos, “Real- time object detection and robotic manipulation for agriculture using a yolo-based learning approach,” 2024

2024
[20]

Spa- tial transformer networks,

M. Jaderberg, K. Simonyan, A. Zisserman, and K. Kavukcuoglu, “Spa- tial transformer networks,” 2016

2016
[21]

Attention mechanisms in computer vision: A survey,

M.-H. Guo, T.-X. Xu, J.-J. Liu, Z.-N. Liu, P.-T. Jiang, T.-J. Mu, S.-H. Zhang, R. R. Martin, M.-M. Cheng, and S.-M. Hu, “Attention mechanisms in computer vision: A survey,”Computational Visual Media, vol. 8, no. 3, p. 331–368, Sep. 2022. [Online]. Available: http://dx.doi.org/...

2022 doi
[22]

Efficient estimation for a smoothing thin plate spline in a two-dimensional space,

J. Cavieres and M. Karkulik, “Efficient estimation for a smoothing thin plate spline in a two-dimensional space,” 2024. [Online]. Available: https://arxiv.org/abs/2404.01902

2024 arXiv
[23]

Enhancing ship classification in optical satellite imagery: Integrating convolutional block attention module with resnet for improved performance,

R. D. Kwon, G. R. Nam, J. Tak, J. Shin, H. Cha, and S. W. Lee, “Enhancing ship classification in optical satellite imagery: Integrating convolutional block attention module with resnet for improved performance,” 2024. [Online]. Available: https: //arxiv.org/abs/2404.02135

2024 arXiv
[24]

Segment anything,

A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y . Lo, P. Doll´ar, and R. Girshick, “Segment anything,” 2023

2023
[25]

Iciou: Improved loss based on complete inter- section over union for bounding box regression,

X. Wang and J. Song, “Iciou: Improved loss based on complete inter- section over union for bounding box regression,”IEEE Access, vol. 9, pp. 105 686–105 695, 2021

2021
[26]

Generalized focal loss: Learning qualified and distributed bounding boxes for dense object detection,

X. Li, W. Wang, L. Wu, S. Chen, X. Hu, J. Li, J. Tang, and J. Yang, “Generalized focal loss: Learning qualified and distributed bounding boxes for dense object detection,”Advances in Neural Information Processing Systems, vol. 33, pp. 21 002–21 012, 2020

2020
[27]

Ultralytics yolov8,

G. Jocher, A. Chaurasia, and J. Qiu, “Ultralytics yolov8,” 2023. [Online]. Available: https://github.com/ultralytics/ultralytics

2023
[28]

Eigen-cam: Class activation map using principal components,

M. B. Muhammad and M. Yeasin, “Eigen-cam: Class activation map using principal components,” in2020 international joint conference on neural networks (IJCNN). IEEE, 2020, pp. 1–7

2020
[29]

melon flower dataset,

Melon, “melon flower dataset,” https://universe.roboflow.com/ melon-oznlh/melon flower, apr 2023, visited on 2024-03-22. [Online]. Available: https://universe.roboflow.com/melon-oznlh/melon flower

2023
[30]

Yolov10: Real-time end-to-end object detection,

A. Wang, H. Chen, L. Liu, K. Chen, Z. Lin, J. Han, and G. Ding, “Yolov10: Real-time end-to-end object detection,”arXiv preprint arXiv:2405.14458, 2024

2024 arXiv
[2004]

Available: https://gfx.cs.princeton.edu/pubs/Brown 2004 NRA/3DPVT04.pdf

[Online]. Available: https://gfx.cs.princeton.edu/pubs/Brown 2004 NRA/3DPVT04.pdf

2004

Pith tools

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

[1] [1]

Yolo deep learning algorithm for object detection in agriculture: a review,

A. Singh and R. Sharma, “Yolo deep learning algorithm for object detection in agriculture: a review,”ResearchGate, 2024. [Online]. Available: https://www.researchgate.net/publication/387051427 YOLO deep learning algorithm for object detection in agriculture a review

work page arXiv 2024

[2] [2]

Advancements in yolo for agricultural object detection,

J. Li and H. Wang, “Advancements in yolo for agricultural object detection,”arXiv, 2024. [Online]. Available: https://arxiv.org/abs/2401. 10379

work page 2024

[3] [3]

Precise Apple Detection and Localization in Orchards using YOLOv5 for Robotic Harvesting Systems

J. Ziyue, Y . Bo, and L. Boyun, “Precise apple detection and localization in orchards using yolov5 for robotic harvesting systems,” 2024. [Online]. Available: https://arxiv.org/abs/2405.06260

work page Pith review arXiv 2024

[4] [4]

Artificial intelligence in agriculture market size report, 2030,

Grand View Research, “Artificial intelligence in agriculture market size report, 2030,” 2023, accessed: 2025-04-02. [On- line]. Available: https://www.grandviewresearch.com/industry-analysis/ artificial-intelligence-in-agriculture-market

work page 2023

[5] [5]

Spatial Transformer Network YOLO Model for Agricultural Object Detection

Y . Zambre, E. Rajkitkul, A. Mohan, and J. Peeples, “Spatial transformer network yolo model for agricultural object detection,” 2024. [Online]. Available: https://arxiv.org/abs/2407.21652

work page Pith review arXiv 2024

[6] [6]

Spatial transformer networks,

M. Jaderberg, K. Simonyan, A. Zisserman, and K. Kavukcuoglu, “Spatial transformer networks,”Advances in Neural Information Processing Systems (NeurIPS), 2015. [Online]. Available: https: //papers.neurips.cc/paper/5854-spatial-transformer-networks.pdf

work page 2015

[7] [7]

Limitations of affine transformations in spatial transformer networks,

R. Brown and Y . Sun, “Limitations of affine transformations in spatial transformer networks,”Princeton Graphics Group,

work page

[8] [8]

Smooth extrapolation of unknown anatomy via statistical shape models,

R. B. Grupp, H. Chiang, Y . Otake, R. J. Murphy, C. R. Gordon, M. Armand, and R. H. Taylor, “Smooth extrapolation of unknown anatomy via statistical shape models,” inMedical Imaging 2015: Image-Guided Procedures, Robotic Interventions, and Modeling, R. J. Webster and Z. R. Yaniv, Eds., vol. 9415. SPIE, Mar. 2015, p. 941524. [Online]. Available: http://dx....

work page doi:10.1117/12.2081310 2015