Image-Based Sorghum Head Counting When You Only Look Once

Eric Hare; Hieu Pham; Lawrence Mosley; Yogesh Bansal

REVIEW 3 major objections 6 minor 17 references

Image-Based Sorghum Head Counting When You Only Look Once

T0 review · 3 major / 6 minor · reviewed 2026-08-27 · deepseek-v4-flash

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

Pith's one-line read A YOLO detector with hand-tuned anchor boxes counts sorghum heads in drone images with out-of-sample mAP 0.95, beating the default model.

desk verdict A plausible sorghum-counting application with a useful anchor-tuning story, but the reported 0.95 mAP is not reproducible as written because the test images are called unlabeled and the baseline test mAP is missing. read the letter →

arxiv 2009.11929 v3 pith:IYDQNKLX submitted 2020-09-24 cs.CV cs.LG

Lawrence Mosley , Hieu Pham , Yogesh Bansal , Eric Hare This is my paper · ORCID

classification cs.CVcs.LG

keywords sorghumheadcountingobjectdetectionYOLOanchorboxtuningaerialdroneimagerymeanaverageprecisiondigitalagriculturephenotyping

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

Sorghum head counting in aerial drone images is a bottleneck in plant breeding because manual counting is slow and expensive. This paper claims that a single-shot YOLOv4 detector, with anchor boxes hand-selected from an exploratory plot of the bounding-box dimensions in the training set, can automatically identify and count sorghum heads. On 300 labeled training images and 50 out-of-sample images, the tuned model reaches a training mAP of 0.99 and a test mAP of 0.95, and the authors report it outperforms the default YOLO model, which uses k-means-derived anchor boxes. The practical promise is that low-cost drone imagery plus a tuned off-the-shelf detector can replace manual phenotyping in commercial breeding programs.

What carries the argument

Anchor boxes are the preset bounding-box shapes that YOLO regresses toward at three up-sampling layers; they define the size distribution the detector can output. Default YOLO generates nine anchors by k-means clustering on the training set's bounding-box coordinates, but the paper shows these cluster centroids do not cover the small heads in the data. The replacement set is 13 anchors: a 10x10 box for the extreme lower tail, plus nine boxes sampled along a linear fit to the height-width scatterplot, placed in regions of above-average variance, and allocated three to the first layer, four to the second, and six to the third. This hand-tuned anchor geometry is the load-bearing mechanism that carries the claimed improvement in mAP and count agreement.

What would settle it

Take the 50 test images, have an independent annotator label every sorghum head as ground truth, and run the parameter-tuned YOLO model on them at IoU 0.70; if the resulting mAP falls materially below 0.95, or if the default YOLO model achieves a higher mAP on the same labels, the central claim is refuted. Also run the default model on those same relabeled test images and report its mAP directly.

Watch

Extended reading notes

Core claim

The paper's central discovery is that the default anchor boxes in YOLO—shapes chosen by k-means clustering over training bounding boxes—miss the small sorghum heads that dominate the lower tail of the size distribution, and that replacing them with an informed anchor set improves detection. By plotting every labeled bounding box's height against its width, the authors saw that the default anchors were too large and poorly placed. They fit a linear model through the scatter, sampled nine anchor boxes along the line, added a 10x10 box for small heads, and distributed 13 anchors across YOLO's three up-sampling layers. With this configuration and an IoU threshold of 0.70, the model scored 0.99 mAP on training images and 0.95 mAP on 50 test images, while also producing a true-versus-predicted count regression with R-squared 0.9513; the default YOLO model's corresponding R-squared was 0.9016. The claim is that accurate sorghum head detection is obtainable through anchor-box tuning with minimal changes to the architecture.

Load-bearing premise

The 0.95 out-of-sample mAP rests on the 50 test images having trustworthy ground-truth bounding boxes even though the data section calls them unlabeled; if those labels are missing, inconsistently produced, or created by a different annotator, the headline accuracy cannot be reproduced or compared.

Editorial extensions

If this is right

A tuned single-shot detector can count sorghum heads from drone images at 0.95 mAP, so manual head-by-head counting in breeding programs could be replaced or prioritized by automated aerial counts.
Because YOLO is a single-shot architecture, the same detector could run near real time on drone-captured video once deployed, making in-season decision support about crop health and yield feasible.
The improvement comes specifically from detecting small heads; models relying on default anchors likely undercount small or distant objects, so anchor coverage should be checked whenever object size varies.
The same exploratory technique—plotting bounding-box height against width before training—can inform anchor choice for other single-class crop counting tasks using aerial imagery.

Reading between the lines

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

The anchor-tuning insight is not specific to sorghum: any single-shot detector applied to small, roughly uniform objects in high-altitude imagery would benefit from checking anchor coverage before training, since the k-means default is driven by object frequency, not by detection difficulty.
A likely extension is to count at multiple IoU thresholds and report count-level error (such as mean absolute count error) alongside mAP, because a farmer cares about total head count more than exact box localization; mAP can be high while count error remains non-negligible.
A direct extension would be to report the default model's test mAP on the same 50 images, which would isolate the anchor-box change as the cause of improvement rather than training duration or random initialization.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Request a human review

A listed scientist reviews the paper for a fee and the review publishes here regardless of verdict. See the reviewers or get listed.

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

Dear Colleague,

Quick take on arXiv:2009.11929: it's a short applied paper on sorghum head counting with YOLOv4, and the interesting bit is the anchor-box tuning story. But the headline 0.95 out-of-sample mAP is not supportable as written: Section 2 says the 50 test images are unlabeled, yet Section 3 reports mAP and true-vs-predicted counts for those same images, which requires ground-truth boxes. That's a load-bearing contradiction, and it has to be fixed before the number means anything.

What's actually useful: the authors do a simple exploratory analysis of the training-set bounding-box sizes and show that the default k-means anchors from YOLO are a poor fit for sorghum heads — they overshoot the large end and miss the small ones. Their hand-tuned anchors, sampled along a linear fit plus a small 10x10 box, are plotted against the data and look sensible. That kind of visual sanity-checking of anchors is practical and could help other small-object detection tasks. The way they assign different numbers of anchors to the three upsampling layers is also a reasonable engineering choice.

The soft spots beyond the test-label contradiction: the default YOLO model only has a training mAP reported (92.0%), so the claim that the tuned model "outperformed the baseline" has no test-set number behind it. The mAP is computed at a fixed IoU of 0.70, which is not a standard evaluation protocol — most detection papers report at 0.50 or 0.50:0.95, and a single non-standard threshold can flatter a model. The test set is only 50 images, so the 0.95 mAP has a nontrivial uncertainty that isn't discussed. No code, weights, or data are released, so the number can't be independently checked. The anchor tuning itself is manual and the authors note it was informed by the training-set distribution; that is fine as long as the test set was held out, but the labeling ambiguity makes that unclear.

Is the central idea wrong? Probably not. The approach is a plausible application of a known detector to a new crop, and the anchor-tuning heuristic is a reasonable extension of common practice. The paper is not methodologically novel, and it doesn't compare to other crop-counting deep learning systems, but it's a legitimate engineering write-up.

Who is this for? Practitioners in digital agriculture who want a worked example of adapting a generic detector to a specific small-object counting problem. A serious referee could get value from it, but only after the authors resolve the test-label contradiction and provide a proper baseline and standard metrics.

My recommendation: if this comes across your desk, don't desk-reject it outright — send it to review with instructions that the evaluation section must be reworked. If the labeling issue is a typo and the test set is actually labeled, the paper could be acceptable as a short application note. As is, the central number is not verifiable.

Best.

Referee Report

3 major / 6 minor

Summary. The paper describes an application of YOLOv4 to count sorghum heads in drone imagery. The authors perform exploratory analysis of bounding-box size distributions and use it to hand-tune a set of 13 anchor boxes. They report a training mAP of 0.99 and an out-of-sample test mAP of 0.95, and they claim the parameter-tuned model outperforms the default YOLO model. The manuscript also reports R² values of 0.9016 (default) and 0.9513 (tuned) for true versus predicted counts at an IoU threshold of 0.70.

Significance. If the reported results are reproducible, the paper offers a practical, low-cost approach to high-throughput sorghum head counting that could be useful in breeding programs. The strength of the work is the explicit data-driven motivation for anchor-box selection, and the use of a real aerial dataset is a relevant testbed. However, the evaluation is incomplete in ways that directly affect the central quantitative claims, so the current manuscript does not firmly establish the 0.95 mAP figure or the claimed superiority over the baseline.

major comments (3)

[Section 2 vs. Section 3] Section 2 states '350 aerial images ... separated into 300 labeled training and 50 unlabeled test images,' yet Section 3 reports a '0.95 mAP on the final test set' (Table 4) and plots 'True Count' versus 'Predicted Count' (Figure 9) with R² values. Computing mAP and true-versus-predicted counts requires ground-truth annotations for the test images. If the test images were labeled after the 300/50 split, the labeling protocol must be disclosed; if they were not labeled, the reported numbers cannot be produced. This contradiction is load-bearing for the headline result and must be resolved.
[Table 4] Table 4 reports only the default YOLO model's training value (92.0%) with no test-set mAP. The abstract and the text claim that the parameter-tuned model 'outperformed the baseline model,' but no test-set comparison with the default model is reported in the table. Since Figure 9 does show a default-model R², the authors clearly have test-set predictions for the default model; they should report the corresponding test mAP so the performance comparison is complete and verifiable.
[Section 3, evaluation protocol] The evaluation protocol is insufficiently specified. The paper reports a single point estimate of 0.95 mAP at a fixed IoU threshold of 0.70, but provides no error bars, no number of repeated runs, no train/validation split, and no training hyperparameters (epochs, learning rate, batch size, early stopping). mAP at a single IoU is non-standard for object detection; it would strengthen the claim to report the IoU threshold explicitly and to provide standard metrics such as mAP@0.5 or mAP@[0.5:0.95] for comparability with the literature.

minor comments (6)

[Throughout] Multiple typographical and language errors appear, such as 'Figuur' (Figures 1, 2, 3, 5, 6, 7), 'Tabel' (Tables 1, 2, 3, 4), and 'Referenties' in the References section. These should be corrected.
[Section 1] The sentence 'Section 3 outlines the methodology of our approach, while results.' is incomplete; it should say 'while Section 3 outlines the methodology and results' or similar.
[Section 2.1] The citation placeholder '(??)' appears in the sentence about deep learning models performing classification; a proper reference should be inserted.
[Table 4] The column heading 'Training Error' is misleading because the values are mAP percentages (92.0%, 99.0%, 95.0%). The heading should be 'Training mAP' and 'Testing mAP' to match the values and avoid confusion.
[Figures 6 and 7] The captions inconsistently refer to 'gray' and 'grey' for the same color. Please unify the spelling.
[Figure 9] The R² values are reported without a definition in the main text. Please specify that R² is the coefficient of determination for the linear regression of predicted counts on true counts, and state how the regression was fit.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the anchor-box tuning is a standard hyperparameter choice informed by training-set statistics and evaluated on held-out test images.

full rationale

The paper's central claim is that a YOLOv4 model with manually tuned anchor boxes achieves an out-of-sample mAP of 0.95 and outperforms a default YOLO baseline. The derivation chain is empirical rather than formal: exploratory analysis of the 300 labeled training images (bounding-box size distributions, counts, coverage) motivates a set of 13 custom anchor boxes placed along a linear fit through training-box dimensions and in the lower tail. These anchors are a hyperparameter configuration, not a fitted predictor of the test labels; the test set is separate. The claimed result, 0.95 test mAP, is an evaluation metric computed against ground truth on that held-out set, and a high training mAP is not claimed as evidence of test performance. No equation is shown to be equivalent to its own input, no parameter is fitted to the test-set outcome and then called a prediction, and no load-bearing argument rests on a self-citation. The cited prior work by the authors (WheatNet) is background literature and is not used to justify the present result. The manuscript has genuine reproducibility and consistency concerns that are not circularity: Section 2 calls the 50 test images 'unlabeled' while Table 4 and Figure 9 report test mAP and true-versus-predicted counts, and the default model's test-set mAP is omitted, weakening the baseline comparison. These are correctness/evidence gaps, not circular derivations, so the circularity score is 0.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central result rests on hand-picked anchor boxes and a chosen IoU threshold, both selected from the training data. The paper introduces no new entities. Transfer from pretrained weights and annotation accuracy are standard domain assumptions.

free parameters (2)

Anchor box dimensions (13 hand-picked boxes) = (10,10),(16,16),(19,19),(16,24),(24,20),(23,24),(28,27),(23,35),(32,32),(38,39),(50,50),(60,60),(80,80)
Selected by hand from a scatterplot of training bounding-box dimensions; the central claim depends on this choice.
IoU detection threshold = 0.70
Set to 0.70 to return all potential heads; the reported mAP is computed at this threshold, so scores are not comparable to standard mAP across IoU thresholds.

assumptions (4)

domain assumption YOLOv4 with Darknet-53 pretrained on ImageNet provides transferable features for sorghum head detection.
The model is initialized with pretrained weights and fine-tuned; this is standard but unproven for this specific domain.
domain assumption The training bounding-box annotations are accurate and complete.
The mAP computation treats these annotations as ground truth; any labeling errors propagate to the reported scores.
ad hoc to paper A line fitted through bounding-box dimensions is a good prior for sampling anchor boxes.
This heuristic is introduced without justification or comparison to alternative anchor selection methods.
domain assumption The test images carry ground-truth labels despite being described as unlabeled.
Section 2 calls the test set unlabeled, but Section 3 uses it for mAP and count comparisons, implying labels exist.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Image-Based Sorghum Head Counting When You Only Look Once." pith.science (2026). https://pith.science/paper/IYDQNKLX

@misc{pith2026200911929,
  author       = {Pith},
  title        = {Pith review of: Image-Based Sorghum Head Counting When You Only Look Once},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IYDQNKLX}},
  note         = {Machine review of arXiv:2009.11929}
}

read the original abstract

Modern trends in digital agriculture have seen a shift towards artificial intelligence for crop quality assessment and yield estimation. In this work, we document how a parameter tuned single-shot object detection algorithm can be used to identify and count sorghum head from aerial drone images. Our approach involves a novel exploratory analysis that identified key structural elements of the sorghum images and motivated the selection of parameter-tuned anchor boxes that contributed significantly to performance. These insights led to the development of a deep learning model that outperformed the baseline model and achieved an out-of-sample mean average precision of 0.95.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

17 extracted references · 14 canonical work pages

[1]

YOLOv4: Optimal Speed and Accuracy of Object Detection

author Bochkovskiy, A. , author Wang, C.Y. , author Liao, H.Y.M. , year 2020 . title Yolov4: Optimal speed and accuracy of object detection . journal arXiv preprint arXiv:2004.10934

work page Pith review arXiv 2020
[2]

, author Lenain , R

author Cariou , C. , author Lenain , R. , author Thuilot , B. , author Martinet , P. , year 2010 . title Autonomous maneuver of a farm vehicle with a trailed implement: motion planner and lateral-longitudinal controllers , in: booktitle 2010 IEEE International Conference on Robotics and Automation , pp. pages 3819--3824 . :10.1109/ROBOT.2010.5509447

work page arXiv 2010
[3]

Blei and John D

author Davis, J. , author Goadrich, M. , year 2006 . title The relationship between precision-recall and roc curves , in: booktitle Proceedings of the 23rd International Conference on Machine Learning , publisher ACM , address New York, NY, USA . pp. pages 233--240 . http://doi.acm.org/10.1145/1143844.1143874, :10.1145/1143844.1143874

work page arXiv 2006
[4]

, author Schmidhalter, U

author Gn \" a dinger, F. , author Schmidhalter, U. , year 2017 . title Digital counts of maize plants by Unmanned Aerial Vehicles (UAVs) . journal Remote Sensing volume 9 , pages 544 . :10.3390/rs9060544

work page doi:10.3390/rs9060544 2017
[5]

, author Safaei, N

author Khaki, S. , author Safaei, N. , author Pham, H. , author Wang, L. , year 2022 . title Wheatnet: A lightweight convolutional neural network for high-throughput image-based wheat head detection and counting . journal Neurocomputing volume 489 , pages 78--89

work page 2022
[6]

, author Lanckriet, G

author McFee, B. , author Lanckriet, G. , year 2010 . title Metric learning to rank , in: booktitle Proceedings of the 27th International Conference on International Conference on Machine Learning , pp. pages 775--782

work page 2010
[7]

, author Malathi, P

author Padala, A.K.R. , author Malathi, P. , year 2022 . title An optimized object detection system using you only look once algorithm and compare with tiny-yolo algorithm with increased accuracy , in: booktitle 2022 2nd International Conference on Innovative Practices in Technology and Management (ICIPTM) , organization IEEE . pp. pages 606--610

work page 2022
[8]

YOLO-LITE: A Real-Time Object Detection Algorithm Optimized for Non-GPU Computers

author Pedoeem, J. , author Huang, R. , year 2018 . title YOLO-LITE: A real-time object detection algorithm optimized for non-gpu computers . journal CoRR volume abs/1811.05588 . http://arxiv.org/abs/1811.05588, http://arxiv.org/abs/1811.05588 arXiv:1811.05588

work page Pith review arXiv 2018

Show all 17 references

[9]

, author Payne, A

author Qureshi, W.S. , author Payne, A. , author Walsh, K.B. , author Linker, R. , author Cohen, O. , author Dailey, M.N. , year 2017 . title Machine vision for counting fruit on mango tree canopies . journal Precision Agriculture volume 18 , pages 224--244 . https://doi.org/1...

2017 doi
[10]

, author Tsoi, N

author Rezatofighi, H. , author Tsoi, N. , author Gwak, J. , author Sadeghian, A. , author Reid, I. , author Savarese, S. , year 2019 . title Generalized intersection over union: A metric and a loss for bounding box regression , in: booktitle The IEEE Conference on Computer Vi...

2019
[11]

, author Rasool, T

author Shaikh, T.A. , author Rasool, T. , author Lone, F.R. , year 2022 . title Towards leveraging the role of machine learning and artificial intelligence in precision agriculture and smart farming . journal Computers and Electronics in Agriculture volume 198 , pages 107119

2022
[12]

, author Langridge, P

author Tester, M. , author Langridge, P. , year 2010 . title Breeding technologies to increase crop production in a changing world . journal Science volume 327 , pages 818--822 . https://science.sciencemag.org/content/327/5967/818, :10.1126/science.1183700, http://arxiv.org/ab...

2010 doi
[13]

, author Balocco, S

author Thomas, G. , author Balocco, S. , author Mann, D. , author Simundsson, A. , author Khorasani, N. , year 2021 . title Intelligent agricultural machinery using deep learning . journal IEEE Instrumentation & Measurement Magazine volume 24 , pages 93--100

2021
[14]

, author Satler , M

author Tripicchio , P. , author Satler , M. , author Dabisias , G. , author Ruffaldi , E. , author Avizzano , C.A. , year 2015 . title Towards smart farming and sustainable agriculture with drones , in: booktitle 2015 International Conference on Intelligent Environments , pp. ...

2015 doi
[15]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...
[16]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...
[17]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 27, 2026 · model on record in the stance chip above.

[1] [1]

YOLOv4: Optimal Speed and Accuracy of Object Detection

author Bochkovskiy, A. , author Wang, C.Y. , author Liao, H.Y.M. , year 2020 . title Yolov4: Optimal speed and accuracy of object detection . journal arXiv preprint arXiv:2004.10934

work page Pith review arXiv 2020

[2] [2]

, author Lenain , R

author Cariou , C. , author Lenain , R. , author Thuilot , B. , author Martinet , P. , year 2010 . title Autonomous maneuver of a farm vehicle with a trailed implement: motion planner and lateral-longitudinal controllers , in: booktitle 2010 IEEE International Conference on Robotics and Automation , pp. pages 3819--3824 . :10.1109/ROBOT.2010.5509447

work page arXiv 2010

[3] [3]

Blei and John D

author Davis, J. , author Goadrich, M. , year 2006 . title The relationship between precision-recall and roc curves , in: booktitle Proceedings of the 23rd International Conference on Machine Learning , publisher ACM , address New York, NY, USA . pp. pages 233--240 . http://doi.acm.org/10.1145/1143844.1143874, :10.1145/1143844.1143874

work page arXiv 2006

[4] [4]

, author Schmidhalter, U

author Gn \" a dinger, F. , author Schmidhalter, U. , year 2017 . title Digital counts of maize plants by Unmanned Aerial Vehicles (UAVs) . journal Remote Sensing volume 9 , pages 544 . :10.3390/rs9060544

work page doi:10.3390/rs9060544 2017

[5] [5]

, author Safaei, N

author Khaki, S. , author Safaei, N. , author Pham, H. , author Wang, L. , year 2022 . title Wheatnet: A lightweight convolutional neural network for high-throughput image-based wheat head detection and counting . journal Neurocomputing volume 489 , pages 78--89

work page 2022

[6] [6]

, author Lanckriet, G

author McFee, B. , author Lanckriet, G. , year 2010 . title Metric learning to rank , in: booktitle Proceedings of the 27th International Conference on International Conference on Machine Learning , pp. pages 775--782

work page 2010

[7] [7]

, author Malathi, P

author Padala, A.K.R. , author Malathi, P. , year 2022 . title An optimized object detection system using you only look once algorithm and compare with tiny-yolo algorithm with increased accuracy , in: booktitle 2022 2nd International Conference on Innovative Practices in Technology and Management (ICIPTM) , organization IEEE . pp. pages 606--610

work page 2022

[8] [8]

YOLO-LITE: A Real-Time Object Detection Algorithm Optimized for Non-GPU Computers

author Pedoeem, J. , author Huang, R. , year 2018 . title YOLO-LITE: A real-time object detection algorithm optimized for non-gpu computers . journal CoRR volume abs/1811.05588 . http://arxiv.org/abs/1811.05588, http://arxiv.org/abs/1811.05588 arXiv:1811.05588

work page Pith review arXiv 2018