PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Therin Young, Elijah Rodriguez, Lisa Coffey, Talukder Zaki Jubery, Adarsh Krishnamurthy, Patrick Schnable, Baskar Ganapathysubramanian

arXiv:2608.30161v1cs.CV

TL;DR

Maize-ear phenotyping needs scalable 3-D measurements because existing methods are constrained by labor, cost, and specialized hardware. The paper develops a low-cost video-to-mesh pipeline using COLMAP, NeRF, automatic fiducial scaling, and quality control, with validation across a diverse ear panel. The platform provides a foundation for breeding-scale 3-D ear phenotyping.

  • Problem

    Manual and existing high-throughput maize-ear phenotyping methods are labor-intensive or constrained by cost and specialized hardware, despite the importance of ear geometry for genetic analyses and yield prediction.

  • Method

    A stationary-camera NeRF pipeline reconstructs a watertight maize-ear mesh from video, using multi-seed COLMAP poses, a known-diameter cylindrical holder for metric scaling, and geometric quality control.

  • Results

    250 of 300 ears (83.3%) passed automated processing and quality control, with skeleton length agreeing with calipers and convex-hull volume agreeing with water displacement.

  • Takeaways & Limitations

    The platform provides a foundation for breeding-scale 3-D maize-ear phenotyping from low-cost video capture with unattended downstream processing.

  • Takeaways & Limitations

    Fifty of 300 ears failed trait extraction, with 35 lost at holder segmentation, and excluded ears were not characterized for possible geometry-related attrition.

Abstract

from arXiv · show

Maize ear geometry (length, width, curvature, and volume) is closely tied to yield and grain-filling outcomes, but existing high-throughput phenotyping pipelines remain constrained by the cost, labor, and specialized hardware they require. We developed and validated a low-cost pipeline that reconstructs a watertight 3-D mesh of a maize ear from a single 20-second video captured with a consumer-grade DSLR on a motorized turntable under uniform LED illumination. Camera poses from a multi-seed COLMAP procedure initialize a Neural Radiance Field (NeRF), and a cylindrical holder of known diameter, visible in every frame, provides automatic metric scaling with downstream geometric quality control. Applied to 300 ears spanning a diverse maize inbred panel, 250 (83.3%) passed automated processing and quality control. Skeleton length agreed with manual caliper measurements across all 250 ears (R^2 = 0.964, RMSE = 4.68 mm), and convex-hull volume agreed with water-displacement volume on a 15-ear subset spanning the full size range (R^2 = 0.982, RMSE = 5.26 mL). Residual length error grew with ear curvature, whereas bounding-box height, which records the same straight-line chord as calipers, showed no such trend; the discrepancy therefore originates in the measurement definition, since calipers record the chord while skeleton length traces the geodesic arc. The capture hardware costs approximately 607 USD, and operator involvement fell from roughly five minutes to one minute per ear, with all downstream processing running unattended. The platform provides a foundation for breeding-scale 3-D ear phenotyping.

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

This platform combines a low-cost stationary-camera NeRF workflow with automated scaling and quality control to recover 3-D maize-ear traits from video. It reports validated length and volume measurements across hundreds of ears.

  • The stationary-camera NeRF pipeline produces 3-D maize-ear traits from video using low-cost hardware.The workflow integrates video capture, COLMAP pose estimation, NeRF reconstruction, scaling, quality control, and trait export.
  • 100% frame registration is enforced by running ten independent COLMAP reconstructions and retaining only full-coverage solutions.Intrinsic parameters are averaged across the retained runs.
  • A known-diameter cylindrical holder provides mesh scaling, while quality control flags failed geometric fits.The holder is visible during capture and supports automatic metric scaling.
  • Length and volume measurements match caliper and water-displacement references across 250 maize ears.The supplied contribution summary reports agreement for both traits across the processed dataset.

1. Introduction

The introduction motivates a breeding-scale 3-D maize-ear phenotyping pipeline by contrasting laborious existing approaches with a frugal workflow that combines NeRF reconstruction, automatic scaling, and quality control.

  • High-resolution maize-ear traits support genetic analyses and yield prediction, but manual phenotyping is labor-intensive, subjective, and difficult to scale.Variation in ear geometry also requires pipelines that capture geometric and physical complexity.
  • Existing approaches span low-cost 2-D methods, classical multi-view stereo, active 3-D sensors, and learning-based reconstruction pipelines.These alternatives differ in geometry recovered, hardware requirements, calibration burden, and throughput.
  • Relative to prior 3-D ear-phenotyping studies, the platform combines stationary-camera NeRF reconstruction with automated scaling and quality control at breeding scale.
  • The proposed pipeline integrates COLMAP pose estimation, Nerfstudio’s nerfacto model, and convex-hull surface reconstruction.It is presented as a frugal and reproducible approach.
  • Ten independent COLMAP runs retain only 100%-coverage solutions and average intrinsics across the retained reconstructions.
  • A known-diameter cylindrical fiducial enables automatic metric scaling without external calibration objects.
  • The workflow wraps all stages from video capture through trait-CSV generation in a single command.

2. Materials and Methods

The platform combines stationary-camera video capture, multi-seed COLMAP and NeRF reconstruction, automatic holder-based metric scaling, geometry quality control, and mesh-derived trait extraction. A low-cost DSLR, turntable, LED lighting, and cylindrical holder support automated per-ear outputs.

  • Trait extraction: The scaled reconstruction becomes a watertight convex-hull mesh, is voxelized and skeletonized, and is divided into ten cross-sections for trait computation.Derived outputs include skeleton length, maximum width and position, chord/arc measures, taper-to-tip, eccentricity, volume, and surface area.
  • Output: The pipeline produces a per-ear trait CSV after automated reconstruction and geometric processing.The algorithmic workflow writes computed size and shape traits to a per-ear CSV; NeRF training requires approximately fourteen minutes per ear on an NVIDIA A100.
  • Data acquisition: The imaging rig uses a consumer-grade DSLR, a motorized turntable rotating at approximately 3 rpm, two off-axis LED panels, a matte-black backdrop, and an 84 mm cylindrical holder.Videos are captured at 30 fps and 1080p for approximately 20 seconds, yielding roughly 600 frames per ear.
  • Experimental material: The study processes a random subset of 300 ears from a 370-genotype diversity panel, with 250 ears completing end-to-end reconstruction.The remaining 50 ears were excluded automatically at COLMAP registration or geometry quality-control stages.
  • 3-D reconstruction: The workflow estimates camera poses with multi-seed COLMAP, trains a nerfacto NeRF, and exports a dense point cloud.Ten independent COLMAP reconstructions vary by random seed; retained camera poses and intrinsics initialize NeRF training.
  • Scaling and quality control: The point cloud is cleaned, isolated to the cylindrical holder, metrically scaled, and screened by automated geometry quality control.Holder-region points are isolated using vertical point-density structure before geometric scaling and downstream processing.

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

The pipeline validates metric scaling by isolating a known cylindrical holder and checking its fitted cross-section before extracting 3-D ear traits. It then cleans, reconstructs, skeletonizes, and measures the ear geometry, including size, curvature, taper, and cross-sectional shape.

  • Quality-control rationale: Holder isolation is verified before trait computation because incorrect boundaries can produce partial-ear, mixed-region, or wrong-region fits and erroneous scale factors.Failure arises when the vertical density profile lacks a clear inflection point, such as with holder occlusion, rim proximity, or an unusually dense ear base.
  • Quality-control indicators: Two indicators detect holder-segmentation failures: fitted eccentricity, where zero denotes a circular holder section, and fitted radius in normalized NeRF units.Radius is expected to remain within a narrow reproducible session range because camera, lens, and camera-to-object distance are held constant.
  • Automatic rejection: Ears are flagged when holder eccentricity exceeds the upper Tukey IQR fence or fitted radius falls outside the expected session range.Visually anomalous holder sections are also excluded and logged with indicator values and exclusion reasons.
  • Trait extraction: The pipeline extracts eleven whole-ear traits and per-section measurements across ten sections, covering size, width, chord-versus-arc straightness, taper, curvature, eccentricity, area, surface area, and volume.Cross-sectional profiles are measured along the skeleton, while convex-hull geometry supplies global surface area and volume.
  • Geometric reconstruction: After cleaning, the point cloud becomes a watertight convex-hull mesh, is voxelized into a solid, and yields a smoothed medial centerline whose arc length defines skeleton length.DBSCAN removes stray fragments, statistical filtering suppresses reconstruction noise, and distance-transform weighting extracts the medial core.
  • Trait interpretation: Chord/Arc Ratio measures axial straightness, whereas Taper-to-Tip is explicitly a shape proxy rather than a direct measure of kernel fill without kernel-level mesh segmentation.The chord-to-arc metric is unavailable from ordinary 2-D projections, while taper cannot distinguish geometric narrowing from tip kernel abortion.

3. Results

The pipeline achieved breeding-scale 3-D maize ear reconstruction and validation across a diverse panel, while reducing operator effort and exposing traits unavailable to manual calipers. Geometric agreement was strong, with curvature-dependent length discrepancies explained by differing measurement definitions.

  • Pipeline performance: 250 / 300 ears (83.3 %) passed automated processing and geometry quality control after COLMAP registration and reconstruction.At least one of ten COLMAP runs fully registered 285 ears; 35 registered ears were later excluded by geometry quality control.
  • Population phenotypes: The reconstructions captured diverse ear size, shape, curvature, and kernel-color phenotypes without manual parameter tuning.Contrasting ears showed coordinated variation in size traits and distinct curvature and taper profiles among shape traits.
  • Geometric validation: R2 = 0.964 and RMSE = 4.68 mm measured agreement between automated skeleton length and manual calipers across 250 ears.Calipers measure the straight-line chord, whereas skeleton length follows the geodesic arc along curved ears.
  • Geometric validation: Skeleton-length error increased with curvature, while bounding-box height error was insensitive to curvature.The relationships were r = −0.54, p < 0.001 for skeleton length and r = −0.05, ns for bounding-box height, indicating a measurement-definition discrepancy rather than reconstruction error.
  • Geometric validation: R2 = 0.982 and RMSE = 5.26 mL measured agreement between convex-hull volume and water-displacement volume across 15 ears spanning the full size range.The convex hull showed a small positive bias because it encloses empty space within curved-ear concavities.
  • Human-effort comparison: Operator involvement fell from approximately five minutes to one minute per ear, while the pipeline produced traits inaccessible to manual caliper measurement.These included convex-hull volume, surface area, cross-sectional profiles, and five shape descriptors; downstream processing ran unattended.

4. Discussion

The platform combines low-cost hardware, automated reconstruction, metric scaling, and quality control to produce validated 3-D maize ear traits at breeding-relevant scale. Its main limitations are holder-segmentation failures, convex-hull volume bias for curved ears, a human visual checkpoint, and limited validation across panels and curvature ranges.

  • Practical impact: $607 hardware, approximately 80% less operator time, and validation against independent references meet three practical breeding-program requirements.The rig uses a consumer DSLR, motorized turntable, LED panels, and a 3-D-printed holder; downstream processing follows a 20-second capture.
  • Technical choices: Ten-seed COLMAP with intrinsic averaging and a fixed-camera rotating-object setup improve pose robustness and simplify capture.Retaining only reconstructions with 100% frame registration safeguards against feature-matching failures, while averaged intrinsics stabilize NeRF initialization.
  • Trait interpretation: Bounding-box height matches calipers because both measure the straight-line chord, whereas skeleton length traces the curved ear’s geodesic arc.Skeletonization of the closed watertight surface is required for skeleton length; curvature therefore predicts divergence from the caliper reference.
  • Biological relevance: Convex-hull volume and surface area correlate with manually measured ear weight, with volume showing the stronger association.The reported associations are R^2 = 0.738 for volume and R^2 = 0.675 for surface area.
  • Limitations: 50 of 300 ears failed trait extraction, with holder segmentation accounting for 35 failures and identifying metric scaling as the least robust stage.Excluded ears can be re-imaged, but the study did not characterize them against retained ears, so geometry-correlated attrition remains unresolved.
  • Limitations: The convex hull overestimates curved-ear volume by spanning the concave flank, limiting agreement with water displacement.A concavity-preserving surface reconstruction is identified as the direct route to improved volumetric accuracy.
  • Limitations and future work: The quality-control system remains semi-automated because visual inspection is needed for a failure mode missed by holder eccentricity.The planned extensions include COLMAP-free pose estimation, but its performance has not yet been formally evaluated.
  • Limitations: Validation used one diversity panel in one season, and the Chord/Arc ratio ranged only from 0.965 to 1.000.Operator and session effects were not isolated, so stronger curvature phenotypes require further evaluation.

5. Conclusions

The study concludes that a low-cost, fixed-camera NeRF pipeline can reconstruct maize ears in 3-D from short videos and extract validated geometric traits. It processed 250 of 300 ears successfully, while scale recovery was the dominant source of attrition.

  • Conclusion: The pipeline reconstructs maize ears from a single 20-second video using a fixed consumer-grade DSLR, multi-seed COLMAP, NeRF, and cylindrical-fiducial scaling.It extracts skeleton length, convex-hull volume, and shape descriptors from watertight meshes.
  • Conclusion: R^2 = 0.964 for skeleton length against manual calipers and R^2 = 0.982 for convex-hull volume against water displacement demonstrate geometric agreement.The corresponding RMSE values are 4.68 mm and 5.26 mL, respectively.
  • Conclusion: 250 of 300 ears, or 83.3%, passed automated processing and quality control, with 35 of 50 exclusions occurring during holder segmentation.Operator involvement was approximately one minute per ear, and downstream processing ran unattended.

CRediT authorship contribution statement

The authorship statement assigns contributions across software, methodology, validation, analysis, investigation, writing, visualization, data curation, conceptualization, and supervision.

  • CRediT authorship contribution statement: Contributors are credited for software, validation, formal analysis, investigation, methodology, writing, visualization, data curation, conceptualization, and supervision.The supplied statement lists individual author roles but is truncated after supervision.

Funding

The work was supported by the AI Institute for Resilient Agriculture and Iowa State University’s Plant Science Institute.

  • Funding: Funding came from the AI Institute for Resilient Agriculture through USDA-NIFA 2021-67021-35329 and Iowa State University’s Plant Science Institute.The funders had no involvement in study design, data work, writing, or the submission decision.

Declaration of generative AI and AI-assisted technologies in the manuscript preparation process

The authors used Claude for language editing and organization during manuscript preparation, then reviewed and edited the content and retained responsibility for the final text.

  • Claude assisted with language editing and manuscript organization.
  • The authors reviewed and edited the AI-assisted content.
  • The authors take full responsibility for the final text.

Geometry quality-control failures

Geometry quality control excluded representative ears with reconstruction failures, including incomplete ear meshes caused by an underestimated holder–ear junction.

  • Figure 14 shows reconstructed holder cross-sections for ears excluded by the geometry quality-control filter.
  • Partial ear detection can underestimate the holder–ear junction and exclude the ear’s upper portion from the segmented region.The resulting ear mesh passed to trait extraction is incomplete.
  • Manual inspection identified partial detection because the holder cross-section alone does not flag it.
Loading 2608.30161v1…