Tracking Moose using Aerial Object Detection

Christopher Indris; Goetz Bramesfeld; Guanghui Wang; Raiyan Rahman

REVIEW 4 major objections 5 minor 62 references

Tracking Moose using Aerial Object Detection

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read For aerial moose detection, a 2.6-million-parameter model matches a 99-million-parameter one — once images are cut into patches.

desk verdict The 93% headline doesn't survive a cross-check against the paper's own per-patch-size table, so the 'lightweight detector is enough' claim is currently unsupported despite a useful comparison framework. read the letter →

arxiv 2507.21256 v1 pith:WSQQ2XJR submitted 2025-07-28 cs.CV

Christopher Indris , Raiyan Rahman , Goetz Bramesfeld , Guanghui Wang This is my paper

classification cs.CV

keywords aerialobjectdetectionwildlifetrackingmoosesurveysmallpatchingaugmentationcomparativestudyUAVdeploymentYOLOv11

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

Moose are counted from crewed aircraft partly because small drones are assumed to lack the onboard compute for a capable detector. This paper tests that assumption by comparing three very different detectors — a 2.6-million-parameter one-stage network, a 64-million-parameter two-stage network, and a 99-million-parameter transformer — on aerial moose imagery cut into large, medium, and small patches. It finds the smallest, fastest model also the most accurate: 93.2% versus 93.1% and 93.0% on the standard mAP@IoU=0.5 score, where a prediction counts only if it overlaps the true moose box by at least half. The paper's conclusion is that for this task the patching recipe, not the architecture, sets the accuracy, and that a lightweight model could plausibly run a drone-based survey — a cheaper, quieter alternative to crewed flights for conservation counts.

What carries the argument

The carrying mechanism is a multi-scale patching augmentation: each full aerial image is divided into large, medium, and small patches whose dimensions are normalized across the four native image resolutions, so the detector sees the same moose at several magnifications. From 1,694 full-size images, the default patching produces 1,771 annotated patches. Two hyperparameters govern the cut: the threshold T, the minimum fraction of a bounding box that must survive for its annotation to be kept (values 0.1, 0.3, 0.5, 1.0), and the overlap O, the fraction by which neighbouring patches overlap (values 0.0, 0.1, 0.3). Patching enlarges moose that would otherwise shrink to a few pixels after a full-image resize, and it exposes the model to partially visible moose at patch borders. The twelve threshold-overlap settings per model, analysed with a cubic-threshold interaction regression whose adjusted R² exceeds 0.96 for both metrics, are what convert the patching recipe into quantitative claims: threshold drives accuracy (T=0.1 is much worse while even 30%-visible moose stay learnable), overlap slightly lowers accuracy by enlarging patches, and model identity barely moves it.

What would settle it

Two checks would settle it. Re-split the released dataset so the validation set's average bounding-box area matches the training set's (about 7,000 pixels rather than the current 10,178), and recompute mAP@IoU=0.5 with the released YOLOv11n weights; a score well below 93.2% would falsify the deployment claim for exactly the harder conditions the training data represent. Or run a field survey: fly the drone at the planned altitude over an area stocked with a known number of moose, count detections, and compare with the actual tally — the gap is the operational accuracy.

Watch

Extended reading notes

Core claim

On its own terms, the paper's claim is that architecture is a weak factor for this detection task once images are patched. YOLOv11n, a one-stage convolutional network with 2.6 million parameters running at 640×640, reaches 93.2% mAP@IoU=0.5; Co-DETR, a transformer-based detector with 99 million parameters running at 800×1200, reaches 93.1%; and Faster R-CNN with 64 million parameters reaches 93.0%. Each model passes the 93% bar on at least one of the twelve threshold-overlap configurations, and their averages across all twelve settings sit within 1.8 points of one another (86.4%–88.2%). The stricter mAP@IoU=0.5:0.95 metric shows the same ordering, with YOLOv11n at 73.3% versus 72.4% for Co-DETR and 71.2% for Faster R-CNN. The regression over all 36 model-threshold-overlap combinations finds small model coefficients, and the paper concludes that faster, simpler models are about as effective as models that require more computational power for this task. The authors draw the practical corollary themselves: for this dataset, model choice matters mainly through speed and memory, so the 5 MB, 6.4 GFLOPs YOLOv11n is the natural fit for a drone, and the patching hyperparameters — not the architecture — are where the accuracy is won or lost.

Load-bearing premise

The load-bearing assumption is that the validation split represents real survey conditions: the validation moose are on average about 45% larger in pixels than the training moose (10,178 vs. 7,019 bounding-box area), so if actual moose are as small as the training ones, operational accuracy will be lower than the reported ~93% even if the model ranking survives.

Editorial extensions

If this is right

A detector the size of YOLOv11n (5.22 MB of weights, 6.4 GFLOPs, about 3.04 GB peak RAM at batch 16) can run detection at 640×640 at roughly five times the per-iteration speed of Co-DETR, making onboard drone surveys a realistic design point.
Even a memory- or speed-limited pipeline that feeds only large patches to the detector keeps strong accuracy, so skipping small patches does not force a collapse in performance.
Because models can learn from moose that are only about 30% visible, patched training handles animals cut by the image boundary — a common event in survey footage.
With the three architectures finishing within 0.2 points of one another at their best settings, deployment choice can be driven by weight, speed, and memory rather than expected accuracy.
The paper identifies expanding the dataset to more varied lighting and designing a patching method that balances positive and negative regions as the next steps, enabled by the planned public release of data and models.

Reading between the lines

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

The reported ~93% scores are likely optimistic for real deployments: the validation split contains moose that are on average about 45% larger in pixels than the training split's (10,178 vs. 7,019 bounding-box area), so small, distant moose are under-represented in the test set and the easy cases are over-represented.
Overlapping patches put the same moose in multiple inference frames, and the paper does not specify how per-patch detections are merged into a final count; an abundance estimate would need cross-patch suppression or voting, and the error of that merger is not measured here.
A natural extension would be to re-split the released dataset so validation matches the training box-size distribution, or to add altitude and lighting strata, then check whether the lightweight model's edge — or its 93% — survives on the hard cases.

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 headline result doesn't hold up against the paper's own Table 7. For T=0.5, O=0.0, Table 5 lists mAP@0.5 = 0.931 for YOLOv11n; Table 7 reports the same configuration broken down by patch size: 0.553, 0.595, and 0.773 on large, medium, and small patches, weighted by boxes to about 0.65. That's not a rounding error. Either the two tables use different evaluation protocols or one set of numbers is wrong. Since the 'lightweight is enough' conclusion is anchored to the Table 5 peaks, the central operational accuracy claim is currently unsupported.

What the paper does well: it's the first direct three-way comparison of YOLOv11n, Faster R-CNN, and Co-DETR for aerial moose detection, and the comparison protocol is sensible. The authors are transparent about the train/validation distribution shift (validation moose are ~45% larger in bbox area), which is a real concern for absolute accuracy. The regression analysis is honest: they fit a cubic threshold term, use cross-validation, and explicitly flag the extrapolation danger.

The soft spots are exactly where the reader and the stress-test point. The threshold hyperparameter changes the test set itself: at T=1.0, all edge-cut moose vanish from the ground truth, and the 93% peaks are all at T=1.0. At T=0.1 the same models are in the 74-79% range. So the headline number is the easiest version of the task. The validation-distribution shift means even the relative ordering might not transfer to real survey conditions. There are single runs only, and the two heavy models get 5 epochs with best-epoch selection, which is a bit thin for a claim of 'as effective as.' Code and data are promised but not released.

The stress-test concern is correct and it is the main problem. Before the 93% claim can be trusted, the authors need to reconcile Table 5 with Table 7. If it's a typo, fine; if not, the conclusion needs to be rewritten around the per-size numbers.

Who this is for: practitioners deciding whether a 2.6M-parameter YOLOv11n can replace a 99M Co-DETR on a drone. The relative ordering (lightweight is competitive) is plausible and probably survives the flaws, but the absolute accuracy is not. I'd send it to peer review because the comparison is useful and the inconsistency is fixable, but I'd require the code/data release and a reconciled set of numbers before accepting it.

Referee Report

4 major / 5 minor

Summary. This paper assembles a moose-detection dataset from four public aerial image sources, applies a multi-scale patching augmentation with two hyperparameters (visibility threshold T and overlap O), and compares YOLOv11n, Faster R-CNN (x101-64x4d), and Co-DETR (R50) under a 12-configuration grid. It reports mAP@IoU=0.5 and mAP@IoU=0.5:0.95 for all 36 model-configuration pairs, fits linear regressions to these results, and adds per-patch-size YOLOv11 experiments. The paper's central claims are that all three detectors reach at least 93% mAP@IoU=0.5 on some configuration and that the much lighter YOLOv11n is about as accurate as the heavier models, making it suitable for UAV deployment.

Significance. If the numbers were internally consistent, the study would be a useful application-oriented benchmark: it compares three architecturally diverse detectors on the same patched dataset, includes per-patch-size analyses, and commits to releasing the dataset and models. The regression diagnostics and explicit discussion of overfitting in Appendix B are positive features. However, the central 93% claim currently rests on Table 5, which is numerically incompatible with the per-patch-size Table 7, and the absolute accuracy is measured on a validation split whose objects are larger than in training. The relative ordering of the three detectors may survive a correction, but the deployment claim does not yet have the needed support.

major comments (4)

[Appendix A (Table 5) and Appendix C (Table 7)] For YOLOv11n at T=0.5, O=0.0, Table 5 reports mAP@IoU=0.5 = 0.931 and mAP@IoU=0.5:0.95 = 0.712, while Table 7 reports per-size mAP@0.5 values of 0.553 (large), 0.595 (medium), and 0.773 (small) on validation subsets with 306, 276, and 351 boxes. A box-count-weighted combination of these per-size APs is about 0.65 for mAP@0.5 and 0.47 for mAP@0.5:0.95. Although AP is not exactly a linear aggregate over subsets, it is implausible that the same evaluation protocol yields an aggregate AP far above the highest subset value for both metrics. The same pattern appears at T=0.3, O=0.0 (Table 5: 0.919 vs. a weighted per-size value around 0.73). Because Table 5 is the source of the 'at least 93%' and 'lightweight is enough' conclusions, the two tables must be reconciled or one of them is wrong.
[Section 3, Table 2 and Appendix C, Table 7] The dataset counts are inconsistent across the two halves of the paper. For the 'default' configuration (T=0.5, O=0.1), Section 3 and Table 2 state that the dataset contains 1,771 patches (1,408 training, 363 validation) after discarding empty patches, while Table 7 rows with the same hyperparameters report 2,275 training images and 606 validation images (summing the large/medium/small columns), with 4,197 training boxes. The annotation counts differ as well. The authors should clarify whether Table 7 includes empty patches, uses a different split, or contains a typo; as written, the training/validation sets used in the two parts of the study cannot be the same.
[Section 3, threshold definition; Table 3] The threshold hyperparameter changes the ground-truth test set: annotations whose visible fraction is below T are discarded before evaluation. The peak values in Table 3 (T=1.0 for all three models) are therefore measured on a validation set from which all edge-cut partial instances have been removed, i.e., the easiest version of the test set. The T=0.1 rows in Table 5, which retain partial instances, are only 74-79% mAP@0.5. The abstract's 'at least 93%' claim should state this qualification, and the conclusion that the models are robust to partial instances should be based on the T=0.1/T=0.3 columns rather than the T=1.0 peaks.
[Table 2, validation distribution] Table 2 shows a sizable train/validation distribution shift: average bounding-box area is 10,178 px in validation vs 7,019 px in training (roughly 45% larger), and normalized area is 0.017 vs 0.013. If the validation moose are systematically closer or larger, the reported absolute mAP values overstate operational accuracy on typical survey imagery. Because all three models are evaluated on the same validation split, the comparative ranking is less affected, but the deployment claim should be hedged or supported by size-stratified results.

minor comments (5)

[Figure 2 caption] The caption states that 'model, threshold and overlap all significantly affect mAP,' but in Fig. 2c the Faster R-CNN coefficient is not statistically significant (p=0.64); only the YOLOv11n coefficient is significant among model terms.
[Table 4] Table 4 should identify the Faster R-CNN backbone (the footnote to Table 5 says x101-64x4d) and should clarify that the reported sec/iter and GFLOPs are measured on an A100 rather than on embedded/edge hardware.
[Appendix B] The phrase 'extrapolate to threshold-overlap value pairs not explicitly tested' should be flagged as extrapolation and accompanied by prediction intervals or a small out-of-sample check, since the 95% bands in Fig. 5 are confidence bands for the mean response, not for new runs.
[Full text] The full text contains numerous spacing and rendering artifacts (e.g., 'challenge s', 'computa tional', 'efﬁciency'); please ensure the final version is free of these.
[Section 5.2] The claim that 'even if the onboard algorithm only passes large patches to the detector ... very strong performance can be attained' is presented without the corresponding Table 7 numbers; at T=0.5, O=0.0, large-patch mAP@0.5 is 0.553, which should be described accurately.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central mAP comparison is a held-out empirical measurement, and the only self-citations are non-load-bearing provenance for the patching method.

full rationale

The paper's central result is an empirical benchmark: the mAP@IoU=0.5 values in Table 5 are measured from trained detectors on a held-out validation split, not quantities derived from the patching hyperparameters or from the regression equations. The linear models in Appendix B (Eqs. 1-9) are descriptive summaries of the 36 measured runs; their coefficients are used only to comment on the measured data and to identify which hyperparameters correlate with mAP, and the paper itself labels out-of-sample use as an 'opportunity to extrapolate' rather than as a validated prediction. The self-citations [20,21] are used only to say the patching approach was 'employed ... in prior works' and to reference edge-cutting behavior; neither citation supplies a theorem, fitted parameter, or uniqueness argument on which the model comparison rests, so it is not load-bearing. The apparent numerical inconsistency between Table 5 (overall mAP 0.931 for YOLOv11n at T=0.5, O=0.0) and Table 7 (per-patch-size mAP@0.5 of 0.553, 0.595, and 0.773 on validation subsets) is a serious correctness or protocol concern, but it is not circularity: no quantity in the paper is defined in terms of the conclusion, and no fitted input is renamed as a prediction. Under the rubric, that concern belongs to correctness risk, not to the circularity score.

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

The paper introduces no new theoretical entities or fitted constants. Its free parameters are experimental design choices (patch scales, threshold, overlap) and the coefficients of descriptive regression models. The central claim relies on the domain assumption that the four public Roboflow datasets provide accurate labels and that the split is representative, both of which are plausible but unverified.

free parameters (4)

Patch scale factors = large/medium/small: 1/6, 1/8, 1/12 for 4000x6000; 1/2, 1/3, 1/4 for 1365x2048 (Table 1)
Hand-selected to make patch dimensions comparable across image sizes; affects object magnification and compute.
Threshold grid values = T in {0.1, 0.3, 0.5, 1.0}
Experimental hyperparameter controlling minimum visible fraction of a bounding box; also changes the test set.
Overlap grid values = O in {0.0, 0.1, 0.3}
Experimental hyperparameter controlling patch overlap and effective patch size.
Linear regression coefficients = 10-19 coefficients depending on model (Eq. 5, Eq. 13)
Estimated from the 36 experimental mAP values; used for statistical commentary and extrapolation, not for the central detection claim.

assumptions (3)

standard math IoU and mAP metrics are computed as defined by standard object-detection practice.
Used throughout Section 5 and Appendix A; no independent verification provided.
domain assumption Annotations in the four Roboflow source datasets are accurate ground truth.
Section 3 assembles the dataset from Refs [53-56] without re-verifying labels.
domain assumption The train/validation split is representative and independent.
Section 3/Table 2 gives counts but not the split mechanism; validation moose are larger on average, so representativeness may fail.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Tracking Moose using Aerial Object Detection." pith.science (2026). https://pith.science/paper/WSQQ2XJR

@misc{pith2026250721256,
  author       = {Pith},
  title        = {Pith review of: Tracking Moose using Aerial Object Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WSQQ2XJR}},
  note         = {Machine review of arXiv:2507.21256}
}

read the original abstract

Aerial wildlife tracking is critical for conservation efforts and relies on detecting small objects on the ground below the aircraft. It presents technical challenges: crewed aircraft are expensive, risky and disruptive; autonomous drones have limited computational capacity for onboard AI systems. Since the objects of interest may appear only a few pixels wide, small object detection is an inherently challenging computer vision subfield compounded by computational efficiency needs. This paper applies a patching augmentation to datasets to study model performance under various settings. A comparative study of three common yet architecturally diverse object detectors is conducted using the data, varying the patching method's hyperparameters against detection accuracy. Each model achieved at least 93\% mAP@IoU=0.5 on at least one patching configuration. Statistical analyses provide an in-depth commentary on the effects of various factors. Analysis also shows that faster, simpler models are about as effective as models that require more computational power for this task and perform well given limited patch scales, encouraging UAV deployment. Datasets and models will be made available via https://github.com/chrisindris/Moose.

Figures

Figures reproduced from arXiv: 2507.21256 by the authors.

**Figure 2.** Summary of experimental results comparing mAP@Io [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

**Figure 3.** Residuals vs. Fitted plot of the four basic linear m [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Diagnostic plots for Eq. 5 (top row) and Eq. 9. 0.70 0.75 0.80 0.85 0.90 0.95 0.25 0.50 0.75 1.00 Threshold Predicted mAP@0.50 Overlap 0 0.1 0.3 0.50 0.55 0.60 0.65 0.70 0.25 0.50 0.75 1.00 Threshold Predicted mAP@0.50.0.95 Overlap 0 0.1 0.3 [PITH_FULL_IMAGE:figures/f…

**Figure 5.** Figure 5: For each Overlap value (0.0, 0.1, 0.3), a plot of mAP [PITH_FULL_IMAGE:figures/full_fig_p015_5.png]

**Figure 6.** Figure 6: Linear fit summary, mAP@IoU=0.50 of the Spline mode [PITH_FULL_IMAGE:figures/full_fig_p018_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

62 extracted references · 39 canonical work pages

[1]

‘It’s like a connection between all of us’: Inuit social connectio ns and caribou declines in Labrador, Canada,

D. Borish, A. Cunsolo, J. Snook, I. Shiwak, M. Wood, A. Dal e, C. Flowers, J. Goudie, A. Hudson, C. Kippenhuck, M. C. Purcell, G. Russell Jr., J. Townley, I. J . Mauro, C. E. Dewey, and S. L. Harper, “‘It’s like a connection between all of us’: Inuit social connectio ns and caribou declines in Labrador, Canada,” Ecology and Society , vol. 27, no. 4, p. 11...

work page doi:10.5751/es-13237-270 2022
[2]

Vulnerability of Inuit food systems to food i nsecurity as a consequence of cli- mate change: a case study from Igloolik, Nunavut,

J. D. Ford, “Vulnerability of Inuit food systems to food i nsecurity as a consequence of cli- mate change: a case study from Igloolik, Nunavut,” Regional Environmental Change , vol. 9, no. 2, pp. 83–100, Jun. 2009. doi: https://doi.org/10.1007 /s10113-008-0060-x. [Online]. Available: https://link.springer.com/article/10.1007/s10113-008-0060-x

work page doi:10.1007/s10113-008-0060-x 2009
[3]

Response of moose to forest ha rvest and management: a literature review,

C. J. Johnson and R. V . Rea, “Response of moose to forest ha rvest and management: a literature review,” Canadian Journal of F orest Research , vol. 54, no. 4, pp. 366–388, Apr. 2024. doi: 10.1139/cjfr-2 023-0158. [Online]. Available: https://cdnsciencepub.com/doi/10.1139/cjfr-2023-015 8

work page doi:10.1139/cjfr-2 2024
[4]

A review of methods to estimate and monitor moose density and abundance,

R. J. Moll, M. K. P . Poisson, D. R. Heit, H. Jones, P . J. Peki ns, and L. Kantar, “A review of methods to estimate and monitor moose density and abundance,” Alces, vol. 58, pp. 31–49, 2022. [Online]. Available: https://alcesjournal.org/index.php/alces/article/view/1881

work page 2022
[5]

A comparison of unmanned aerial vehicles (drones) and manned helicopters for monitoring macropod populations,

M. Gentle, N. Finch, J. Speed, and A. Pople, “A comparison of unmanned aerial vehicles (drones) and manned helicopters for monitoring macropod populations,” Wildlife Research, vol. 45, no. 7, p. 586, 2018. doi: 10.1071/WR18034. [Online]. Available: http://www.publish.csiro.au/?paper=WR18034

work page doi:10.1071/wr18034 2018
[6]

Small-object detection for uav-based images using a distance metric method,

H. Zhou, A. Ma, Y . Niu, and Z. Ma, “Small-object detection for uav-based images using a distance metric method,” Drones, vol. 6, no. 10, p. 308, 2022. doi: https://doi.org/10.3390 /drones6100308. [Online]. Available: https://www.mdpi.com/2504-446X/6/10/308

work page 2022
[7]

Deep learning wor kﬂow to support in-ﬂight processing of digital aerial imagery for wildlife population surveys,

T.-W . Ke, S. X. Y u, M. D. Koneff, D. L. Fronczak, L. J. Fara, T. J. Harrison, K. L. Landolt, E. J. Hlavacek, B. R. Lubinski, and T. P . White, “Deep learning wor kﬂow to support in-ﬂight processing of digital aerial imagery for wildlife population surveys,” PLOS ONE , vol. 19, no. 4, p. e0288121, Apr. 2024. doi: 10.1371/journal.pone.0288121. [Online]. A...

work page doi:10.1371/journal.pone.0288121 2024
[8]

A survey of small object detection bas ed on deep learning in aerial images,

W . Hua and Q. Chen, “A survey of small object detection bas ed on deep learning in aerial images,” The Artiﬁcial intelligence review, vol. 58, no. 6, p. 162, Mar. 2025. doi: 10.1007/s10462-025- 11150-9. [Online]. Available: https://link.springer.com/article/10.1007/s10462-025-11150-9

doi:10.1007/s10462-025- 2025

Show all 62 references

[9]

Small object bird detection in infrared drone videos using mask r- cnn deep learning,

Y . M. Kassim, M. E. Byrne, C. Burch, K. Mote, J. Hardin, D. R . Larsen, and K. Palaniappan, “Small object bird detection in infrared drone videos using mask r- cnn deep learning,” Electronic Imaging , vol. 8, pp. 85–1–85–8, Jan. 2020. doi: 10.2352/ISSN.2470-1 173.2020.8.IMA W...

2020
[10]

Prescribed grass ﬁre mapping and rate of spread measurement using nir images from a small ﬁxed-wing uas,

S. Gowravaram, H. Chao, Z. Lin, S. Parsons, T. Zhao, M. Xi n, X. Hu, P . Tian, H. P . Flanagan, and G. Wang, “Prescribed grass ﬁre mapping and rate of spread measurement using nir images from a small ﬁxed-wing uas,” IEEE Journal of Selected T opics in Applied Earth Observatio ...

2023
[11]

Microsoft COCO: Common Objects in Context,

T.-Y . Lin, M. Maire, S. Belongie, L. Bourdev, R. Girshic k, J. Hays, P . Perona, D. Ramanan, C. L. Zitnick, and P . Dollár, “Microsoft COCO: Common Objects in Context,” Feb. 2015, arXiv:1405.0312 [cs]. [Online]. Available: http://arxiv.org/abs/1405.0312

2015 arXiv
[12]

The Pascal Visual Object Classes (VOC) Challenge,

M. Everingham, L. V an Gool, C. K. I. Williams, J. Winn, an d A. Zisserman, “The Pascal Visual Object Classes (VOC) Challenge,” International Journal of Computer Vision , vol. 88, no. 2, pp. 303–338, Jun. 2010. doi: 10.1007/s11263-009-0275-4. [Online]. Available: http://link.s...

2010 doi
[13]

Visual tra cking of small animals in cluttered natural environments using a freely moving camera,

B. Risse, M. Mangan, L. Del Pero, and B. Webb, “Visual tra cking of small animals in cluttered natural environments using a freely moving camera,” in Proceedings of the IEEE international conference on computer vision workshops , Oct. 2017. doi: 10.1109/ICCVW .2017.335 pp. 284...

2017
[14]

Wildlife monitoring with drones: A survey of end users,

R. B. Iglay, L. R. Jones, J. A. Elmore, K. O. Evans, S. Sami appan, M. B. Pfeiffer, and B. F. Blackwell, “Wildlife monitoring with drones: A survey of end users,” Wildlife Society Bulletin , vol. 48, no. 3, p. e1533, Jun. 2024. doi: 10.1002/wsb.1533. [Online]. Available: https...

2024 doi
[16]

Au tomated detection of wildlife using drones: Synthesis, opportunities and constraints,

E. Corcoran, M. Winsen, A. Sudholz, and G. Hamilton, “Au tomated detection of wildlife using drones: Synthesis, opportunities and constraints,” Methods in ecology and evolution , vol. 12, no. 6, pp. 1103–1114, 2021. doi: 10.1111/2041-210X .13581. [Online]. Available: https://b...

2021 doi
[17]

Ultralytics YOLO,

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

2023
[18]

Faster R-CNN: Tow ards Real-Time Object Detection with Region Proposal Networks,

S. Ren, K. He, R. Girshick, and J. Sun, “Faster R-CNN: Tow ards Real-Time Object Detection with Region Proposal Networks,” IEEE Transactions on Pattern Analysis and Machine Intellig ence, vol. 39, no. 6, pp. 1137–1149, Jun. 2017. doi: 10.1109/TP AMI .2016.2577031. [Online]. Av...

2017
[19]

DETRs with Collaborative H ybrid Assignments Training,

Z. Zong, G. Song, and Y . Liu, “DETRs with Collaborative H ybrid Assignments Training,” in 2023 IEEE/CVF International Conference on Computer Vision (ICCV). Paris, France: IEEE, Oct

2023
[20]

A new dataset and comparative study for aphid cluster dete ction and segmentation in sorghum ﬁelds,

R. Rahman, C. Indris, G. Bramesfeld, T. Zhang, K. Li, X. C hen, I. Grijalva, B. McCornack, D. Flippo, A. Sharda et al. , “A new dataset and comparative study for aphid cluster dete ction and segmentation in sorghum ﬁelds,” Journal of Imaging , vol. 10, no. 5, p. 114, May 2024....

2024 doi
[21]

IEEE, 2023

On the Real-Time Semantic Segmentation of Aphid Clusters in the Wild . IEEE, 2023. doi: 10.1109/CVPRW59228.2023.00670. [Online]. Available: https://ieeexplore.ieee.org/document/10208615

2023
[22]

Y ou Only Look Once: Uniﬁed, Real-Time Object Detection,

J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, “Y ou Only Look Once: Uniﬁed, Real-Time Object Detection,” May 2016, arXiv:1506.02640 [cs]. [Online]. Av ailable: http://arxiv.org/abs/1506.02640

2016 arXiv
[23]

Improvements in Aerial Object Detection: C omparing YOLOv7 with YOLOv5 for Fine Drone and Bird Detection in V olatile Environments,

S. Ramesh, “Improvements in Aerial Object Detection: C omparing YOLOv7 with YOLOv5 for Fine Drone and Bird Detection in V olatile Environments,” Master’s the sis, Dublin, National College of Ireland, Sep. 2023. [Online]. Available: https://norma.ncirl.ie/7256/

2023
[24]

R ecent Advances for Aerial Object Detection: A Survey,

J. Leng, Y . Y e, M. Mo, C. Gao, J. Gan, B. Xiao, and X. Gao, “R ecent Advances for Aerial Object Detection: A Survey,” ACM Comput. Surv. , vol. 56, no. 12, pp. 296:1–296:36, Jul. 2024. doi: 10.1145/ 3664598. [Online]. Available: https://doi.org/10.1145/3664598

2024 doi
[25]

Research Towards Y olo-Series Algorithms: Co mparison and Analysis of Object Detection Models for Real-Time UA V Applications,

S. Wang, “Research Towards Y olo-Series Algorithms: Co mparison and Analysis of Object Detection Models for Real-Time UA V Applications,” Journal of Physics: Conference Series , vol. 1948, no. 1, p. 012021, Jun. 2021. doi: 10.1088/1742-6596/1948/1/012021 Publisher: IOP Publish...

1948 doi
[26]

Comparison of YOL O V ersions for Object Detection from Aerial Images,

M. E. Atik, Z. Duran, and R. Özgünlük, “Comparison of YOL O V ersions for Object Detection from Aerial Images,” International Journal of Environment and Geoinformatics , vol. 9, no. 2, pp. 87–93, Jun. 2022. doi: 10.30897/ijegeo.1010741. [Online]. Available: http://ijetms.in/V ...

2022 doi
[27]

Efﬁcient-L ightweight YOLO: Improving Small Object Detection in YOLO for Aerial Images,

M. Hu, Z. Li, J. Y u, X. Wan, H. Tan, and Z. Lin, “Efﬁcient-L ightweight YOLO: Improving Small Object Detection in YOLO for Aerial Images,” Sensors, vol. 23, no. 14, p. 6423, Jul. 2023. doi: 10.3390/s23146423 . [Online]. Available: https://www.mdpi.com/1424-8220/23/14/6423

2023 doi
[28]

Efﬁcie nt golf ball detection and tracking based on convolutional neural networks and kalman ﬁlter,

T. Zhang, X. Zhang, Y . Y ang, Z. Wang, and G. Wang, “Efﬁcie nt golf ball detection and tracking based on convolutional neural networks and kalman ﬁlter,” arXiv preprint arXiv:2012.09393 , 2020. doi: 10.48550/arXiv.2012.09393. [Online]. Available: https://arxiv.org/abs/2012.09393

2012 doi arXiv
[29]

YOLO-U A V: Object Detection Method of Unmanned Aerial V ehicle Imagery Based on Efﬁcient Multi- Scale Feature Fusion,

C. Ma, Y . Fu, D. Wang, R. Guo, X. Zhao, and J. Fang, “YOLO-U A V: Object Detection Method of Unmanned Aerial V ehicle Imagery Based on Efﬁcient Multi- Scale Feature Fusion,” IEEE Ac- cess, vol. 11, pp. 126 857–126 878, 2023. doi: 10.1109/ACCESS.20 23.3329713. [Online]. Availa...

2023
[30]

MBSDet: A Novel Method for Marin e Object Detection in Aerial Imagery with Complex Background Suppression,

G. Ji, L. Fan, and C. Li, “MBSDet: A Novel Method for Marin e Object Detection in Aerial Imagery with Complex Background Suppression,” Electronics, vol. 13, no. 23, p. 4764, Dec. 2024. doi: 10.3390/electronics13234764. [Online]. Available: https://www.mdpi.com/2079-9292/13/23/4764

2024 doi
[31]

Oriented Bounding Box Representation Based on Continuous Encoding in Oriented SAR Ship Detection,

P . Li, C. Feng, W . Feng, and X. Hu, “Oriented Bounding Box Representation Based on Continuous Encoding in Oriented SAR Ship Detection,” IEEE Journal of Selected T opics in Applied Earth Observatio ns and Remote Sensing , pp. 1–14, 2025. doi: 10.1109/JSTARS.2025.3541217 Confe...

2025
[32]

Aerial Imaging-Based Soiling Detection System for Solar Photovoltaic Panel Cleanliness Inspectio n,

U. Naeem, K. Chadda, S. V ahaji, J. Ahmad, X. Li, and E. Asa di, “Aerial Imaging-Based Soiling Detection System for Solar Photovoltaic Panel Cleanliness Inspectio n,” Sensors, vol. 25, no. 3, p. 738, Jan. 2025. doi: 10.3390/s25030738. [Online]. Available: https://www.mdpi.com/...

2025 doi
[33]

CPDD: A Cross-Scenario Pho tovoltaic Defect Detector Based on Fine-Grained Feature Autoencoding and Pseudo-Box Contr astive Learning,

Z. Wang, H. Chen, and Z. Cao, “CPDD: A Cross-Scenario Pho tovoltaic Defect Detector Based on Fine-Grained Feature Autoencoding and Pseudo-Box Contr astive Learning,” IEEE Transactions on Semiconductor Manufacturing , pp. 1–1, 2025. doi: 10.1109/TSM.2025.3570323. [Online]. Avai...

2025
[34]

Beyond sRGB: Optimizing O bject Detection with Diverse Color Spaces for Precise Wildﬁre Risk Assessment,

Z. Y ang, S. Cao, and M. Aibin, “Beyond sRGB: Optimizing O bject Detection with Diverse Color Spaces for Precise Wildﬁre Risk Assessment,” Remote Sensing, vol. 17, no. 9, p. 1503, Apr. 2025. doi: 10.3390/rs17091503 . [Online]. Available: https://www.mdpi.com/2072-4292/17/9/1503

2025 doi
[35]

Real-Time Aerial Multispectral O bject Detection with Dynamic Modality-Balanced Pixel-Level Fusion,

Z. Wang and Q. Zhang, “Real-Time Aerial Multispectral O bject Detection with Dynamic Modality-Balanced Pixel-Level Fusion,” Sensors, vol. 25, no. 10, p. 3039, May 2025. doi: 10.3390/s25103039. [Online]. Available: https://www.mdpi.com/1424-8220/25/10/3039

2025 doi
[36]

Mask R-C NN,

K. He, G. Gkioxari, P . Dollar, and R. Girshick, “Mask R-C NN,” in Proceedings of the IEEE International Conference on Computer Vision (ICCV) , Oct 2017. doi: 10.1109/iccv.2017.322 pp. 2980–2988. [Onl ine]. Available: https://openaccess.thecvf.com/content_iccv_2017/html/He_Mas...

2017 doi
[37]

IEEE, 2019

Libra R-CNN: T owards Balanced Learning for Object Detection. IEEE, 2019. doi: 10.1109/CVPR.2019.00091. [Online]. Available: https://ieeexplore.ieee.org/document/8953703

2019
[38]

F ocal Loss for Dense Object Detection,

T.-Y . Lin, P . Goyal, R. Girshick, K. He, and P . Dollár, “F ocal Loss for Dense Object Detection,” Feb. 2018, arXiv:1708.02002 [cs]. [Online]. Available: http://arxiv.org/abs/1708.02002

2018 arXiv
[39]

Multispecies detection and identiﬁcation of African mammals in aerial imagery using convolutional ne ural networks,

A. Delplanque, S. Foucher, P . Lejeune, J. Linchant, and J. Théau, “Multispecies detection and identiﬁcation of African mammals in aerial imagery using convolutional ne ural networks,” Remote Sensing in Ecology and Conservation , vol. 8, no. 2, pp. 166–179, Apr. 2022. doi: 10....

2022 doi
[40]

Optimized faster R-CNN for oil wells detection from high-resolution remote s ensing images,

Z. Wang, L. Bai, G. Song, Y . Zhang, M. Zhu, M. Zhao, L. Chen , and M. Wang, “Optimized faster R-CNN for oil wells detection from high-resolution remote s ensing images,” International Journal of Remote Sensing, vol. 44, no. 22, pp. 6897–6928, Nov. 2023. doi: 10.1080/014 31161...

2023
[41]

An I mproved Deep Learning Approach for Retrieving Outfalls Into Rivers From UAS Image ry,

Y . Huang, C. Wu, H. Y ang, H. Zhu, M. Chen, and J. Y ang, “An I mproved Deep Learning Approach for Retrieving Outfalls Into Rivers From UAS Image ry,” IEEE Transactions on Geoscience and Remote Sensing , vol. 60, pp. 1–14, 2022. doi: 10.1109/TGRS.2021.3113901. [Online]. Avail...

2022
[42]

Analysis of the performance of Faster R-CNN and YOLOv8 in de tecting ﬁshing vessels and ﬁshes in real time,

L. Ezzeddini, J. Ktari, T. Frikha, N. Alsharabi, A. Alay ba, A. J. Alzahrani, A. Jadi, A. Alkholidi, and H. Hamam, “Analysis of the performance of Faster R-CNN and YOLOv8 in de tecting ﬁshing vessels and ﬁshes in real time,” PeerJ Computer Science , vol. 10, p. e2033, May 2024...

2024 doi
[43]

Accumulated trivial attention matters in vision transformers on small datasets,

X. Chen, Q. Hu, K. Li, C. Zhong, and G. Wang, “Accumulated trivial attention matters in vision transformers on small datasets,” in Proceedings of the IEEE/CVF winter conference on applicati ons of computer vision , 2023. doi: 10.1109/W ACV56688.2023.00397 pp. 3984–3992. [Onlin...

2023
[44]

Deformabl e DETR: Deformable Transformers for End-to-End Object Detection,

X. Zhu, W . Su, L. Lu, B. Li, X. Wang, and J. Dai, “Deformabl e DETR: Deformable Transformers for End-to-End Object Detection,” Mar. 2021, arXiv:2010.04159 [cs]. [Onl ine]. Available: http://arxiv.org/abs/2010.04159

2021 arXiv
[45]

Depth-wise convolu tions in vision transformers for efﬁcient training on small datasets,

T. Zhang, W . Xu, B. Luo, and G. Wang, “Depth-wise convolu tions in vision transformers for efﬁcient training on small datasets,” Neurocomputing, vol. 617, p. 128998, 2025. doi: 10.1016/j.neucom.2024.12 8998. [Online]. Available: https://arxiv.org/abs/2407.19394 10

2025 arXiv
[46]

A Lightweight CNN–Trans former Network With Laplacian Loss for Low-Altitude UA V Imagery Semantic Segmentation,

W . Lu, Z. Zhang, and M. Nguyen, “A Lightweight CNN–Trans former Network With Laplacian Loss for Low-Altitude UA V Imagery Semantic Segmentation,” IEEE Transactions on Geoscience and Remote Sensing , vol. 62, pp. 1–20, 2024. doi: 10.1109/TGRS.2024.3385318. [Online]. Available:...

2024
[47]

[Online]

Papers with Code - DETRs with Collaborative Hybrid Assi gnments Training. [Online]. Available: https://paperswithcode.com/paper/detrs-with-collabo rative-hybrid-assignments
[48]

Dcef 2-yolo: Aerial detection yolo with deformable convolution–efﬁcient feature fusion for small target dete ction,

Y . Shin, H. Shin, J. Ok, M. Back, J. Y oun, and S. Kim, “Dcef 2-yolo: Aerial detection yolo with deformable convolution–efﬁcient feature fusion for small target dete ction,” Remote Sensing , vol. 16, no. 6, p. 1071, 2024. doi: https://doi.org/10.3390/rs16061071. [Online]. Ava...

2024 doi
[49]

ESOD: Efﬁcient Small Object Detection on High-Resolution Images,

K. Liu, Z. Fu, S. Jin, Z. Chen, F. Zhou, R. Jiang, Y . Chen, a nd J. Y e, “ESOD: Efﬁcient Small Object Detection on High-Resolution Images,” Dec. 2024, ar Xiv:2407.16424 [cs]. [Online]. Available: http://arxiv.org/abs/2407.16424

2024 arXiv
[50]

D etection and tracking meet drones challenge,

P . Zhu, L. Wen, D. Du, X. Bian, H. Fan, Q. Hu, and H. Ling, “D etection and tracking meet drones challenge,” IEEE Transactions on Pattern Analysis and Machine Intellig ence, vol. 44, no. 11, pp. 7380–7399, 2021. doi: 10.1109/TP AMI.2021.3119563. [Online]. Available:https://ie...

2021
[51]

Patch-Level Augmentation for Object Detection in Aerial Images,

S. Hong, S. Kang, and D. Cho, “Patch-Level Augmentation for Object Detection in Aerial Images,” in 2019 IEEE/CVF International Conference on Computer Vision W ork shop (ICCVW) . Seoul, Korea (South): IEEE, Oct. 2019. doi: 10.1109/ICCVW .2019.00021. ISBN 978-1-728 15-023-9 pp. ...

2019
[52]

YOLOv11: An Overview of the Ke y Architectural Enhancements,

R. Khanam and M. Hussain, “YOLOv11: An Overview of the Ke y Architectural Enhancements,” Oct. 2024. doi: 10.48550/arXiv.2410.17725 ArXiv:2410.17725 [cs]. [ Online]. Available: http://arxiv.org/abs/2410.17725

2024 doi arXiv
[53]

moose detect Dataset,

silentwings, “moose detect Dataset,” Jun. 2023. [Onli ne]. Available: https://universe.roboﬂow.com/silentwings/moose-dete ct

2023
[54]

moose Dataset ,

——, “ moose Dataset ,” Jun. 2023. [Online]. Available: https://universe.roboﬂow.com/silentwings/moose-5wmr w

2023
[55]

moose Dataset ,

——, “ moose Dataset ,” Jun. 2023. [Online]. Available: https://universe.roboﬂow.com/silentwings/moose-0wmb y

2023
[56]

Moose detection Dataset ,

——, “ Moose detection Dataset ,” Jun. 2023. [Online]. Av ailable: https://universe.roboﬂow.com/silent-wings/moose-det ection

2023
[57]

YOLOv4: Optimal Speed and Accuracy of Object Detection,

A. Bochkovskiy, C.-Y . Wang, and H.-Y . M. Liao, “YOLOv4: Optimal Speed and Accuracy of Object Detection,” Apr. 2020, arXiv:2004.10934 [cs, eess]. [Online]. Availab le: http://arxiv.org/abs/2004.10934

2020 arXiv
[58]

Eval uating the Evolution of YOLO (Y ou Only Look Once) Models: A Comprehensive Benchmark Study of YOLO1 1 and Its Predecessors,

N. Jegham, C. Y . Koh, M. Abdelatti, and A. Hendawi, “Eval uating the Evolution of YOLO (Y ou Only Look Once) Models: A Comprehensive Benchmark Study of YOLO1 1 and Its Predecessors,” Oct. 2024, arXiv:2411.00201 [cs]. [Online]. Available: http://arxiv.org/abs/2411.00201

2024 arXiv
[59]

Integrating remote sensing and deep le arning into aerial survey of large african mammals,

A. Delplanque, “Integrating remote sensing and deep le arning into aerial survey of large african mammals,” Ph.D. dissertation, ProQuest Dissertations & Theses. ISBN 9798346503965 2024. [Online]. Available: https://www.proquest.com/docview/3132863750

2024
[60]

End-to-End Object Detection with Transformers,

N. Carion, F. Massa, G. Synnaeve, N. Usunier, A. Kirillo v, and S. Zagoruyko, “End-to-End Object Detection with Transformers,” May 2020, arXiv:2005.12872 [cs]. [Onl ine]. Available: http://arxiv.org/abs/2005.12872

2020 arXiv
[61]

OpenMMLab Detection Tool box and Benchmark,

MMDetection Contributors, “OpenMMLab Detection Tool box and Benchmark,” Aug. 2018. [Online]. Available: https://github.com/open-mmlab/mmdetection

2018
[62]

Deep Residual Learni ng for Image Recognition,

K. He, X. Zhang, S. Ren, and J. Sun, “Deep Residual Learni ng for Image Recognition,” in 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Las V egas, NV , USA: IEEE, Jun. 2016. doi: 10.1109/CVPR.2016.90. ISBN 978-1-4673-88 51-1 pp. 770–778. [Online]. Ava...

2016
[2023]

ISBN 979835030 7184 pp

doi: 10.1109/ICCV51070.2023.00621. ISBN 979835030 7184 pp. 6725–6735. [Online]. Available: https://ieeexplore.ieee.org/document/10376521/

2023

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

‘It’s like a connection between all of us’: Inuit social connectio ns and caribou declines in Labrador, Canada,

D. Borish, A. Cunsolo, J. Snook, I. Shiwak, M. Wood, A. Dal e, C. Flowers, J. Goudie, A. Hudson, C. Kippenhuck, M. C. Purcell, G. Russell Jr., J. Townley, I. J . Mauro, C. E. Dewey, and S. L. Harper, “‘It’s like a connection between all of us’: Inuit social connectio ns and caribou declines in Labrador, Canada,” Ecology and Society , vol. 27, no. 4, p. 11...

work page doi:10.5751/es-13237-270 2022

[2] [2]

Vulnerability of Inuit food systems to food i nsecurity as a consequence of cli- mate change: a case study from Igloolik, Nunavut,

J. D. Ford, “Vulnerability of Inuit food systems to food i nsecurity as a consequence of cli- mate change: a case study from Igloolik, Nunavut,” Regional Environmental Change , vol. 9, no. 2, pp. 83–100, Jun. 2009. doi: https://doi.org/10.1007 /s10113-008-0060-x. [Online]. Available: https://link.springer.com/article/10.1007/s10113-008-0060-x

work page doi:10.1007/s10113-008-0060-x 2009

[3] [3]

Response of moose to forest ha rvest and management: a literature review,

C. J. Johnson and R. V . Rea, “Response of moose to forest ha rvest and management: a literature review,” Canadian Journal of F orest Research , vol. 54, no. 4, pp. 366–388, Apr. 2024. doi: 10.1139/cjfr-2 023-0158. [Online]. Available: https://cdnsciencepub.com/doi/10.1139/cjfr-2023-015 8

work page doi:10.1139/cjfr-2 2024

[4] [4]

A review of methods to estimate and monitor moose density and abundance,

R. J. Moll, M. K. P . Poisson, D. R. Heit, H. Jones, P . J. Peki ns, and L. Kantar, “A review of methods to estimate and monitor moose density and abundance,” Alces, vol. 58, pp. 31–49, 2022. [Online]. Available: https://alcesjournal.org/index.php/alces/article/view/1881

work page 2022

[5] [5]

A comparison of unmanned aerial vehicles (drones) and manned helicopters for monitoring macropod populations,

M. Gentle, N. Finch, J. Speed, and A. Pople, “A comparison of unmanned aerial vehicles (drones) and manned helicopters for monitoring macropod populations,” Wildlife Research, vol. 45, no. 7, p. 586, 2018. doi: 10.1071/WR18034. [Online]. Available: http://www.publish.csiro.au/?paper=WR18034

work page doi:10.1071/wr18034 2018

[6] [6]

Small-object detection for uav-based images using a distance metric method,

H. Zhou, A. Ma, Y . Niu, and Z. Ma, “Small-object detection for uav-based images using a distance metric method,” Drones, vol. 6, no. 10, p. 308, 2022. doi: https://doi.org/10.3390 /drones6100308. [Online]. Available: https://www.mdpi.com/2504-446X/6/10/308

work page 2022

[7] [7]

Deep learning wor kﬂow to support in-ﬂight processing of digital aerial imagery for wildlife population surveys,

T.-W . Ke, S. X. Y u, M. D. Koneff, D. L. Fronczak, L. J. Fara, T. J. Harrison, K. L. Landolt, E. J. Hlavacek, B. R. Lubinski, and T. P . White, “Deep learning wor kﬂow to support in-ﬂight processing of digital aerial imagery for wildlife population surveys,” PLOS ONE , vol. 19, no. 4, p. e0288121, Apr. 2024. doi: 10.1371/journal.pone.0288121. [Online]. A...

work page doi:10.1371/journal.pone.0288121 2024

[8] [8]

A survey of small object detection bas ed on deep learning in aerial images,

W . Hua and Q. Chen, “A survey of small object detection bas ed on deep learning in aerial images,” The Artiﬁcial intelligence review, vol. 58, no. 6, p. 162, Mar. 2025. doi: 10.1007/s10462-025- 11150-9. [Online]. Available: https://link.springer.com/article/10.1007/s10462-025-11150-9

doi:10.1007/s10462-025- 2025