AI-enabled Low-Cost 3D Maize Ear Morphometry Platform at Breeding Scale

Highlights

• The low-cost stationary-camera NeRF pipeline derives 3-D maize ear traits from video.Source: Highlights · p. 2
• Ten-seed COLMAP with intrinsic averaging achieves 100% frame registration.Source: Highlights · p. 2
• A holder of known diameter scales the mesh, while quality control flags failed fits.Source: Highlights · p. 2
• A one-command workflow runs frame processing, COLMAP, NeRF, scaling, and trait output to CSV.Source: Highlights · p. 2
• Length and volume match calipers and displacement across 250 maize ears.The supplied highlight does not specify the number of ears used for each comparison.Source: Highlights · p. 2

Introduction

• Manual maize-ear trait characterization is labor-intensive, subjective, and difficult to scale, while 2-D projections cannot recover back-side kernels, true volumes, or complete 3-D spatial distributions.Full-surface geometry enables volumetric traits and cross-sectional profiles unavailable from a single projection.Source: §1 · pp. 3–4
• Classical MVS requires controlled views, careful backgrounds, and expert tuning, while active 3-D sensors are costly, calibration-heavy, and sensitive to reflectance and lighting.Turntable rigs using consumer hardware have made metric 3-D reconstruction of individual fruits more accessible.Source: §1 · p. 4
• The current NeRF mesh supports structurally accurate macroscopic ear traits, not direct 3-D kernel-level measurements; companion study extends the platform to kernel-level extraction.The companion work uses cylindrical point-cloud unwrapping and zero-shot instance segmentation to recover ear- and kernel-level traits.Source: §1 · p. 4
• The pipeline is positioned as the first breeding-scale maize-ear method combining stationary-camera NeRF reconstruction, consumer-grade hardware, automatic metric scaling, and quality control.Table 1 situates it among existing automated maize-ear phenotyping methods.Source: §1 · p. 5, Table 1 · p. 5
• Its design combines ten independent COLMAP reconstructions retaining 100%-coverage solutions, known-diameter cylindrical metric scaling, and a single-command video-to-trait-CSV workflow.Intrinsics are averaged across the retained COLMAP runs.Source: §1 · p. 5

2. Materials and Methods

• The workflow estimates camera poses, trains a NeRF, cleans and scales the resulting point cloud, applies geometry quality control, and produces per-ear traits.The output is a watertight convex-hull mesh that is skeletonized and reduced to the trait set.Source: §2 · pp. 5–6, Figure 1 · p. 6

2.1. Imaging hardware and cost breakdown

• The imaging rig uses a consumer DSLR, motorized turntable, LED panels, and an 84 mm cylindrical holder, with all imaging components costing around 607 USD.The holder is 3-D printed in PLA and serves as the metric fiducial.Source: §2.1 · p. 6, Figure 2 · p. 7

2.2. Data acquisition

• From 300 randomly selected ears in a 370-genotype diversity panel, 250 completed reconstruction; five operators collected data across sessions using turntables rotating at approximately 3 rpm.Each synchronized capture recorded a full 360° rotation at 30 fps and 1080p for approximately 20 seconds, then extracted frames at 3 fps for pose estimation.Source: §2.2 · pp. 6–7

2.3. 3-D reconstruction

• The reconstruction stages include pose estimation, NeRF-based volumetric reconstruction, ear segmentation and metric scaling, then trait extraction from the scaled watertight mesh.The workflow voxelizes and skeletonizes the mesh and divides it into ten cross-sections to compute size and shape traits for a per-ear CSV.Source: Figure 1 · p. 6, §2.3 · p. 7

2.3.1. Pose estimation: multi-seed COLMAP

• Ten independent COLMAP runs use different random seeds; runs below 100% frame-registration coverage are discarded, and the lowest-reprojection-error retained run supplies camera poses.Intrinsics are averaged across retained runs, with focal-length variation below 0.5% relative change in practice.Source: §2.3.1 · p. 8

2.4. Neural radiance field reconstruction

• The nerfacto model trains for 30,000 iterations on an NVIDIA A100 and requires approximately fourteen minutes per ear.It uses camera poses and intrinsics from COLMAP and represents a continuous radiance field with a multi-resolution hash-grid MLP.Source: §2.4 · p. 8

2.5. Metric scaling using the cylindrical holder

• The 84 mm holder diameter sets metric scale: a circle fit to isolated holder points estimates its diameter, and uniform coordinate scaling converts the reconstruction to millimeters.The holder is visible in every frame, avoiding a separate calibration target; averaged COLMAP intrinsics are used to mitigate systematic distortion.Source: §2.5 · p. 9

2.6. Geometry quality control via cylindrical holder cross-section analysis

• Holder-isolation errors can misplace the cutoff and produce partial-ear, ear-plus-holder, or wrong-region fits, so eccentricity and fitted radius serve as automatic checks.A zero eccentricity indicates a circular section; a radius outside the reproducible session range suggests the fit latched onto the wrong point-cloud region.Source: §2.6 · p. 9
• Ears are flagged when eccentricity exceeds the upper Tukey IQR fence or radius falls outside the batch’s expected range; visually anomalous fits are also excluded and logged.The rejection log records eccentricity, radius, and the reason for exclusion.Source: §2.6 · p. 10

2.7. Trait extraction

• The pipeline reports eleven whole-ear traits—six size and five shape measures—plus values at ten cross-sections, following EarCV and EARBOX trait vocabulary.Size traits include skeleton length, maximum width, maximum-width position, convex-hull surface area and volume, and bounding-box height; shape traits include straightness, taper, bending, and roundness.Source: §2.7 · p. 10
• DBSCAN clustering, statistical outlier filtering, and two-pass HSV filtering remove fragments, reconstruction noise, and turntable reflections before Open3D convex-hull meshing yields a watertight surface.DBSCAN uses ε = 5 mm and minimum cluster size 50; the statistical filter uses k = 20 neighbors and a 2σ threshold.Source: §2.7 · p. 11
• A 3-mm voxelized solid and distance-transform medial core produce a smoothed centerline whose arc length defines skeleton length.The medial core is formed from weighted voxel centroids retained slice by slice along the longitudinal axis.Source: §2.7 · pp. 11–12
• PCA aligns the longitudinal axis to z; endpoint chord, convex-hull surface area and volume, maximum width, and its position along the skeleton complement arc length.Maximum width is the largest major diameter across the ten cross-sections.Source: §2.7 · p. 12
• Ten equal-length skeleton sections yield major and minor diameters, eccentricity, and area; Chord/Arc Ratio quantifies 3-D straightness, while Taper-to-Tip is a shape proxy, not a direct kernel-fill measure.Curvature Score normalizes accumulated tangent turning by arc length, and Max Bending captures the sharpest local bend; whole-ear and section-wise traits are saved per ear in a CSV file.Source: §2.7 · p. 12

3.1. COLMAP registration and geometry quality control

• 285 ears (95%) achieved 100% frame registration in at least one of ten COLMAP runs; after geometry quality control, 250/300 ears (83.3%) passed to trait extraction.The other 15 ears required re-capture, and 35 registered ears were excluded by the geometry filter; selecting the lowest-reprojection-error fully registered run supported stable NeRF training.Source: §3.1 · p. 13

3.3. Geometric validation

• R^2 = 0.964 and RMSE = 4.68 mm for skeleton length (n = 250); R^2 = 0.982 and RMSE = 5.26 mL for convex-hull volume (n = 15) against manual references.The volume subset spanned the full size range. Convex-hull volume showed a small positive bias because the hull encloses empty space within curved ears’ concavities.Source: Figure 8 · p. 14, §3.3 · pp. 15–16
• r = −0.54 (p < 0.001) for skeleton-length deviation versus Chord/Arc ratio, while bounding-box-height deviation was insensitive to curvature (r = −0.05, ns).Calipers and bounding-box height measure the straight-line chord, whereas skeleton length traces the longer geodesic arc; the discrepancy reflects measurement definition rather than reconstruction error.Source: Figure 9 · p. 15, §3.3 · p. 15

3.2. Diverse Maize Ears (Qualitative Examples)

• Reconstructions captured diverse ear size, shape, curvature, and kernel color without manual parameter tuning, including short, long, straight, curved, yellow, and red-purple examples.The curated qualitative examples were separate from the 300 ears used in quantitative experiments.Source: §3.2 · p. 15

3.4. Population-level trait distributions

• Population examples showed that width and length can vary semi-independently, while curvature and tapering represent distinct shape dimensions.Across 250 ears, skeleton length ranged from approximately 75 to 225 mm, and Chord/Arc ratio clustered near unity (0.965–1.000). Surface-area and volume associations with ear weight had R^2 = 0.675 (RMSE = 18.70 cm2) and R^2 = 0.738 (RMSE = 17.41 cm3), respectively.Source: §3.4 · p. 16, Figure 10 · p. 17

3.5. Human-effort comparison

• Operator involvement fell from approximately five minutes per ear for manual phenotyping to approximately one minute for the pipeline, an approximately 80% reduction.Downstream processing ran unattended, and the pipeline delivered traits inaccessible to manual caliper measurement, including volume, surface area, cross-sectional profiles, and five shape descriptors.Source: §3.5 · pp. 16–17, Table 5 · p. 17

4.1. Practical impact

• The $607 rig and one-minute operator involvement after a 20-second capture support medium-scale use, though larger-scale throughput and performance across environments remain unevaluated.The workflow runs autonomously on a GPU node, but a visual inspection checkpoint is required for one failure mode.Source: §4.1 · p. 17, §4.4 · p. 19

4.2. Design advantages

• A stationary camera and single-axis rotation simplify synchronization and calibration, while ten seeded COLMAP runs and averaging successful intrinsics guard against feature-matching failures.These design choices may improve consistency, but fixed-camera capture was not formally compared with moving-camera capture.Source: §4.2 · p. 18
• Bounding-box height measures the caliper-like chord, whereas skeleton length follows the curved ear’s geodesic arc; the watertight mesh also enables volume, surface-area, and cross-sectional estimates.Convex-hull volume agreed with water displacement at R2 = 0.982 and RMSE = 5.26 mL, with largest deviations for curved ears. Figure 11 illustrates curvature and Taper-to-Tip varying semi-independently of other shape descriptors.Source: §4.2 · p. 18, Figure 11 · p. 18

4.4. Limitations

• 50 of 300 ears failed first-pass trait extraction, including 35 lost at holder segmentation; re-imaging may reduce routine losses, but excluded ears were not compared with retained ones.Metric scaling is the least robust stage; a re-capture takes a 20-second video and approximately one minute of operator time.Source: §4.4 · p. 19
• Volume estimates are biased upward for curved ears, one quality-control failure mode still requires visual inspection, and evaluation used one panel in one season with a narrow Chord/Arc ratio range.Five operators and two turntable models were used, but operator and session effects were not isolated; the Chord/Arc ratio ranged from 0.965 to 1.000, warranting evaluation on more curved germplasm. Concavity-preserving reconstruction is proposed to reduce volume bias.Source: §4.4 · p. 19

4.5. Future Directions

• Planned extensions include turntable-informed camera-pose initialization refined with NeRF training and kernel-level trait extraction from the dense color point cloud.The kernel pipeline described in a companion study extracts kernel row number, kernels per row, and individual kernel dimensions without crop-specific model training.Source: §4.5 · pp. 19–20

4.3. Biological relevance

• Convex-hull volume and surface area associated with manually measured ear weight at R2 = 0.738 and 0.675, respectively.The corresponding RMSE values were 17.41 cm3 for volume and 18.70 cm2 for surface area. Weight variation also reflects kernel moisture, maturity, and cob-to-kernel ratio; the convex hull overestimates true volume for curved ears, and current mesh resolution does not resolve individual kernels.Source: §4.3 · pp. 18–19, Figure 13 · p. 20

5. Conclusions

• Skeleton length matched manual calipers (R2 = 0.964, RMSE = 4.68 mm), and convex-hull volume matched water-displacement volume (R2 = 0.982, RMSE = 5.26 mL).Both comparisons used the reconstructed watertight meshes.Source: §5 · p. 20
• 250 of 300 SAM diversity-panel ears (83.3 %) passed automated processing and quality control; holder-segmentation quality control excluded 35 ears, versus 15 failures at COLMAP registration.Scale recovery, rather than pose estimation, was the dominant source of attrition.Source: §5 · p. 20
• Operator involvement is approximately one minute per ear, with all downstream processing running unattended.The pipeline combines multi-seed COLMAP pose estimation, a standard nerfacto NeRF model, and cylindrical-fiducial scaling to produce the reconstructions.Source: §5 · p. 20
• Convex-hull volume and surface area were moderately associated with manually measured ear weight (R2 = 0.738 and 0.675, respectively).Residual variance was attributed to kernel moisture content, maturity state, and the convex-hull estimator's geometric limitation for curved ears.Source: §5 · p. 20
• The reported trait set is deliberately ear-level; a companion study extends the platform to eleven ear- and kernel-level traits, while COLMAP-free pose estimation remains under evaluation.Its performance relative to the current pipeline has yet to be formally established.Source: §5 · p. 20

CRediT authorship contribution statement

• Therin Young and Talukder Zaki Jubery contributed software, validation, formal analysis, investigation, visualization, and both original-draft and review-and-editing writing.Jubery also contributed methodology.Source: CRediT authorship contribution statement · p. 21
• Elijah Rodriguez contributed methodology, software, validation, visualization, and review-and-editing writing; Lisa Coffey contributed data curation and review-and-editing writing.Source: CRediT authorship contribution statement · p. 21
• Adarsh Krishnamurthy contributed conceptualization, supervision, funding acquisition, and review-and-editing writing.Source: CRediT authorship contribution statement · p. 21
• Patrick Schnable and Baskar Ganapathysubramanian contributed conceptualization, resources, supervision, project administration, funding acquisition, and review-and-editing writing.Source: CRediT authorship contribution statement · p. 21

Funding

• Support came from the AI Institute for Resilient Agriculture (USDA-NIFA 2021-67021-35329) and Iowa State University’s Plant Science Institute, with no funder involvement in study design, data work, report writing, or the publication decision.Source: Funding · p. 21

• The authors used Claude for language editing and organization, then reviewed and edited the content and took full responsibility for the final text.Source: Declaration of generative AI and… · p. 21

Geometry quality-control failures

• A contaminated segmentation includes ear structure in the fitted holder region, producing a non-circular cross-section and an unreliable scale factor.The eccentricity check detects this failure mode.Source: Figure 14 · p. 24
• Holder mis-identification fits a region other than the cylindrical holder, producing a radius outside the expected session range and an erroneous scale factor.The radius check detects this failure mode before scale errors propagate into physical-unit traits.Source: Figure 14 · p. 24
• Partial ear detection places the cutoff below the holder–ear junction, excluding the upper ear and leaving an incomplete mesh for trait extraction.The holder cross-section can remain approximately circular, so eholder alone does not flag this failure.Source: Figure 14 · p. 24
• Eccentricity and radius checks automatically detect contaminated segmentation and holder mis-identification, respectively.These checks correspond to Eqs. 3–4.Source: Figure 14 · p. 24
• Manual review identified partial ear detection, motivating routine visual inspection of the segmented point cloud.This failure is not flagged by eholder alone.Source: Figure 14 · p. 24
• The figure notes flat z-density profiles, partial holder occlusion, and ears positioned too close to the holder rim as underlying causes across the illustrated cases.Figure 14 shows reconstructed holder cross-sections for representative ears excluded by the geometry quality-control filter.Source: Figure 14 · p. 24, Geometry quality-control failures · p. 24