Corn Ear Detection and Orientation Estimation Using Deep Learning

John Evans; Michael Mardikes; Nathan Sprague

REVIEW 3 major objections 5 minor 1 references

Corn Ear Detection and Orientation Estimation Using Deep Learning

T0 review · 3 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read A robot-mounted camera trained with YOLOv8 and keypoints can estimate 3D corn ear angles with about 18 degrees of mean error, approaching the 12-degree agreement between two humans.

desk verdict A plausible first proof-of-concept for automated 3D corn ear orientation, but the headline 18° MAE is a best-case number that depends on hand-fixed tracking and an unvalidated linear angle model. read the letter →

arxiv 2412.14954 v1 pith:SHLBARYF submitted 2024-12-19 cs.CV cs.LG

Nathan Sprague , John Evans , Michael Mardikes This is my paper · ORCID

classification cs.CVcs.LG

keywords cornearorientationestimationobjectdetectionkeypointmulti-viewtriangulationYOLOv8plantphenotypingfieldrobotics

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 tries to show that a ground robot carrying an RGB camera and running deep learning can replace hand measurements of corn ear orientation. It combines YOLOv8 ear detection, keypoint-based 2D pose, and BoT-SORT tracking to turn a video pass-by into a 3D estimate of each ear's cardinal angle and its angle off the stalk. The reported detector finds about 90 percent of ears, the 2D pose has about 3 degrees mean absolute error, and the final 3D cardinal angle lands within approximately 18 degrees of hand measurements, close to the 12-degree spread between two people measuring the same ears. The accuracy is achieved after tracker identity errors are manually corrected, so the paper's central feasibility claim is about detection, pose, and angle reconstruction, with tracking flagged as the remaining bottleneck.

What carries the argument

The load-bearing object is the line of potential 3D angles in a polar plot, one line per video frame. Each line is fixed by the 2D pose angle from keypoints and by the camera-relative angle derived from the bounding box x-position divided by the field of view; the intersection point of lines from multiple frames encodes the ear's full 3D angle. Supporting machinery is YOLOv8 for ear detection, a keypoint network for the tip and node, and BoT-SORT for maintaining ear identity across hundreds of frames.

What would settle it

Record a calibrated checkerboard or rod of known 3D orientation while driving the robot past it, and compare the pipeline's angle output against the known geometry; if the error grows with distance from image center or with ear curvature, the x-position-to-angle model is biased. Alternatively, re-run the 100-ear evaluation without manually fixing tracker IDs and check whether the MAE stays near 18 degrees or rises above the human-human 12-degree difference.

Watch

Extended reading notes

Core claim

The central discovery is that a monocular video pass-by contains enough angular information to recover a corn ear's 3D orientation without depth sensors. Each frame supplies two angles: the ear's camera-relative direction, read off the bounding box x-coordinate divided by the camera field of view, and the ear's 2D pose angle from keypoints at tip and node. Plotting the possible 3D angles from two or more frames as lines on a polar diagram, the intersection of those lines gives the cardinal angle (the angle of the intersection point about the center) and the angle off the stalk (the point's radial distance). With roughly 260 observations per ear as the robot moves by, the intersection becomes well-constrained, yielding a mean absolute error of approximately 18 degrees with r2=0.943 on a 100-ear test segment, compared with 12 degrees between two human measurers.

Load-bearing premise

The whole reconstruction rests on the assumption that each ear is a straight, rigid segment whose camera-relative angle is obtained by dividing its bounding box x-coordinate by the camera field of view, ignoring perspective projection, lens distortion, and the ear's bending and twisting.

Editorial extensions

If this is right

Ear-drop over time, the increasing angle off the stalk as plants dry down, can be quantified as a continuous curve from repeated robot passes without touching the plants.
Seed-placement experiments can be scaled up: if seed orientation controls leaf orientation and ears originate from leaves, field-wide ear-angle distributions become measurable.
The same angular-resolution pipeline can be pointed at other elongated plant organs and at the harvestability question of how ear angle affects combine damage.
Because accuracy is assessed under manually corrected identities, improving tracking under full occlusion is the next gating step for a fully automated system.

Reading between the lines

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

The 18-degree figure is likely a lower bound on the current method's field accuracy, since it is computed under perfect tracking; if tracker fixes are removed, error on ears that suffer identity swaps will be higher.
The geometric conversion from bounding box x-coordinate to angle by simple division by field of view ignores perspective and lens distortion; calibrating the camera would likely narrow the gap toward the 12-degree human baseline.
The line-intersection method is a generic multi-view rigid-segment pose estimator; it should transfer to stems, stalks, fruit pedicels, or any near-straight organ observed from a moving camera.
A synthetic test with a camera model and rigid rods of known 3D orientation could isolate the geometric bias from the detection and pose errors.

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

Short version: this is a plausible first shot at automated 3D corn ear orientation from RGB video, and the detection/tracking analysis is honest. But the headline 18° MAE is a best-case number: it assumes hand-fixed tracking, and the angle conversion from image x to camera-relative angle is a linear division by FOV, with no calibration or undistortion. Treat the accuracy claim as setup-specific.

What's new: the task itself. Prior work did ear height (Wong) and leaf angle (Xiang), not ear orientation. The pipeline of YOLOv8 detection, SAM-generated segmentation, keypoint pose, BoT-SORT tracking, and line-intersection triangulation is component-wise standard, but the integration for this phenotyping question is new, and the temporal ear-drop study is a nice demonstration.

What's good: the object detection numbers are credible (mAP 0.894), the 2D keypoint pose has MAE 3°, and they explicitly report tracking error rates (36 errors over 2669 observations, affecting 26/100 ears) rather than hiding them. The decision to use keypoints over segmentation because it was more accurate is sensible. The paper is candid that tracking is the bottleneck.

Where it's soft: the geometric reconstruction in §2.3 is the load-bearing part, and it's the least defended. The x-to-angle mapping is stated as 'dividing by the field of view,' which is only a small-angle approximation; for a pinhole camera it should be atan((x-cx)/fx). With cameras at 120–150 cm and wide FOV, ears off-axis will have systematic bias. No calibration, no undistortion, no synthetic validation. The straight-rigid-ear assumption is also acknowledged to be violated (bending, twisting, droop). Second, the 18° MAE is computed after manually fixing tracker ID errors—so it's an upper bound on what the automated system can do in practice. Third, no error bars on the 18° vs 12° comparison; with only 100 ears, that difference may not be significant. Fourth, no code or data released, which makes reproduction hard.

None of this kills the paper. The temporal ear-drop example shows the pipeline produces sensible biological signal despite the geometric simplifications, and the limitations are stated. But the headline accuracy should be reframed as 'feasibility under labor-corrected tracking with a specific camera geometry,' not a validated general accuracy claim.

For whom: plant phenotyping researchers who want a starting point for ear orientation measurement, and vision folks interested in agricultural applications. It deserves a serious referee—the task is new and the pipeline is a reasonable baseline—but a referee should push for calibration, error bars, and a non-oracle tracking evaluation.

Recommendation: send to peer review, with a request for major revision on the geometry and evaluation. I'd bring it to a reading group to discuss what counts as validation in agri-vision.

Referee Report

3 major / 5 minor

Summary. The paper proposes a computer-vision pipeline for detecting corn ears in robot-acquired field video, estimating their 2D pose with keypoints or fitted ellipses, tracking them with BoT-SORT, and combining multi-view 2D angle observations into a 3D ear orientation estimate. The authors report a detection mAP of 0.894, a 2D keypoint pose MAE of 3 degrees, and a 3D cardinal-angle MAE of approximately 18 degrees on 100 hand-measured ears, compared with a 12-15 degree human inter-rater difference, and conclude that automated ear orientation monitoring is feasible.

Significance. If the 3D accuracy claim held, the system would be a useful high-throughput phenotyping tool for maize ear orientation, which is currently measured manually. The object detection and 2D keypoint results are credible and represent real engineering progress. However, the headline 3D accuracy rests on an uncalibrated projective mapping and on hand-corrected tracking; with the evidence as presented, the contribution is a feasible but setup-specific demonstration rather than a validated general method.

major comments (3)

[§2.3] The mapping from the bounding-box x-coordinate to the camera-relative angle θ1 is stated as 'dividing it by the field of view of the camera,' but for a pinhole camera the correct relation is θ1 = atan((x - cx)/fx), and the linear approximation holds only near the optical axis. Since every multi-view intersection in Figure 8 and the resulting 18° MAE in §3.4 is built from θ1, this uncalibrated mapping is load-bearing; the paper reports no camera calibration, no undistortion step, and no synthetic validation, so the headline accuracy cannot be assumed to transfer to other camera heights, lenses, or off-axis ear positions. This needs a derivation or a calibrated replacement, plus a sensitivity analysis.
[§3.3 and §3.4] The automated pipeline produced tracking errors in 26 of 100 ears (§3.3), yet the 3D accuracy evaluation in §3.4 assumes 'perfect tracking' by manually fixing incorrect IDs. The claim that the system can track and predict ear orientation is therefore not supported for the fully automated pipeline; the authors should report the 3D MAE under the actual BoT-SORT tracker output and, if hand correction is retained, present it explicitly as an upper bound under idealized tracking.
[§2.2 and §3.4] The paper states in §2.2 that once ears point downward the hand-measured cardinal angle is only accurate to ±45°, and §3.5 shows that the dataset includes many downward-pointing ears by late season. Including such ears in the 18° MAE comparison against a 12° human inter-rater difference conflates sensor error with an unreliable ground-truth label; the analysis should either exclude or separately analyze downward ears, and the paper should explain why the abstract reports a 15° human inter-rater MAE while §3.4 reports 12°.

minor comments (5)

[§2.3 and §2.3 (object detection)] There are two sections numbered 2.3; the 3D pose estimation section and the object detection section should be renumbered.
[Figure 8] The caption mentions 'Probability lines' but the text describes deterministic lines of possible angles; please clarify the axes, units, and how the green, red, and pink lines are generated.
[§3.1] The comparison of the ear-detector mAP (0.894) with YOLOv9's COCO AP is not informative because the tasks, classes, and evaluation protocols differ; either remove it or provide a matched comparison.
[§2.1] The statement that video was recorded at 720p is not accompanied by timing or inference speed measurements, so the claim of real-time performance is not verified.
[References] Some references have incomplete author lists or formatting errors (e.g., 'Varghese & M. 2024' and the duplicated corn ear damage citations); please correct them.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the 3D angle estimates are triangulated from camera geometry and keypoints, then compared against independent angle-finder measurements.

full rationale

The paper's central claim—automated corn ear orientation estimates reach approximately 18° MAE—is not derived from the ground-truth angles by construction. Section 2.3 obtains θ1 from the bounding-box x-coordinate and camera field of view, and θ2 from keypoints; multi-view line intersections yield the 3D angle. Section 3.4 compares the result to independent digital-angle-finder measurements, with no angle parameter fitted to those measurements. The choice of keypoint pose over segmentation pose (Section 3.2) is test-set model selection, which is a mild statistical selection risk, but it is not a fitted input renamed as a prediction and does not force the reported 3D error. All load-bearing method components (YOLOv8, BoT-SORT, SAM) are cited to external groups; no self-citation chain supports the result. The linear x-to-FOV mapping in Section 2.3 is geometrically approximate and could limit generalization, but that is an accuracy and validity concern, not circularity, because the mapping is not defined in terms of the output angles.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The method has no fitted physical parameters; its accuracy rests on a set of domain assumptions about ear rigidity, camera motion, and ground-truth reliability, plus two un-tuned detection thresholds. No new entities such as particles, forces, or conserved quantities are introduced.

free parameters (2)

detection confidence threshold = 0.25 (default)
Used in the final pipeline in Section 3.1; kept at the YOLOv8 default rather than tuned, but it sets the recall/precision operating point and therefore affects downstream angle estimates.
minimum IOU for matching detections to ground truth = 0.5
Used to evaluate 2D pose accuracy in Section 3.2; this scoring choice affects which detections are counted as correct but is not a model parameter.

assumptions (5)

domain assumption The x-coordinate of a detection maps linearly to the camera-relative angle by dividing by the camera field of view.
Invoked in Section 2.3 to define theta1; this ignores nonlinear perspective projection and camera intrinsics or distortion.
domain assumption A corn ear is straight and rigid between tip and node, so its 2D projected angle across frames traces a line whose intersection yields the true 3D orientation.
Basis of the triangulation shown in Figure 8; the paper itself notes torsional flexibility of stalks and downward bending of ears, which violate this.
domain assumption The camera travels parallel to the row at a fixed height with no roll, and ears are stationary during the pass.
Required for the angle-from-x-coordinate and line intersection geometry in Section 2.3.
domain assumption Two-person digital angle finder measurements are a usable ground truth despite stated uncertainty for drooping ears.
Used as ground truth in Sections 2.2 and 3.4; the paper states hand measurement is prone to human error and can be off by about 45 degrees when ears point straight down.
domain assumption SAM-generated segmentation masks, prompted by bounding boxes, provide training targets accurate enough for the segmentation and pose model.
Section 2.3 describes generating segmentation training images with Segment Anything Model instead of manual labeling, without a reported verification of mask accuracy.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Corn Ear Detection and Orientation Estimation Using Deep Learning." pith.science (2026). https://pith.science/paper/SHLBARYF

@misc{pith2026241214954,
  author       = {Pith},
  title        = {Pith review of: Corn Ear Detection and Orientation Estimation Using Deep Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SHLBARYF}},
  note         = {Machine review of arXiv:2412.14954}
}

read the original abstract

Monitoring growth behavior of maize plants such as the development of ears can give key insights into the plant's health and development. Traditionally, the measurement of the angle of ears is performed manually, which can be time-consuming and prone to human error. To address these challenges, this paper presents a computer vision-based system for detecting and tracking ears of corn in an image sequence. The proposed system could accurately detect, track, and predict the ear's orientation, which can be useful in monitoring their growth behavior. This can significantly save time compared to manual measurement and enables additional areas of ear orientation research and potential increase in efficiencies for maize production. Using an object detector with keypoint detection, the algorithm proposed could detect 90 percent of all ears. The cardinal estimation had a mean absolute error (MAE) of 18 degrees, compared to a mean 15 degree difference between two people measuring by hand. These results demonstrate the feasibility of using computer vision techniques for monitoring maize growth and can lead to further research in this area.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

1 extracted references · 1 linked inside Pith

[1]

Ali,R.M.,&Buchele,W.F.(n.d.).Cornearorientationaffectsmechanicaldamageandforces onconcave

Aharon,N.,Orfaig,R.,&Bobrovsky,B.-Z.(2022).BoT-SORT:RobustAssociations Multi-PedestrianTracking.Retrievedfromhttp://arxiv.org/abs/2206.14651. Ali,R.M.,&Buchele,W.F.(n.d.).Cornearorientationaffectsmechanicaldamageandforces onconcave. Transactionsof theASAE,18(3),0444–0447.doi:10.13031/2013.36607. Fortin,M.-C.,&Pierce,F.J.(1996).LeafAzimuthinStrip-Intercrop...

arXiv 2022

Pith tools

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

[1] [1]

Ali,R.M.,&Buchele,W.F.(n.d.).Cornearorientationaffectsmechanicaldamageandforces onconcave

Aharon,N.,Orfaig,R.,&Bobrovsky,B.-Z.(2022).BoT-SORT:RobustAssociations Multi-PedestrianTracking.Retrievedfromhttp://arxiv.org/abs/2206.14651. Ali,R.M.,&Buchele,W.F.(n.d.).Cornearorientationaffectsmechanicaldamageandforces onconcave. Transactionsof theASAE,18(3),0444–0447.doi:10.13031/2013.36607. Fortin,M.-C.,&Pierce,F.J.(1996).LeafAzimuthinStrip-Intercrop...

arXiv 2022