MaizeHorizon: a forward-motion, range-stratified per-plant maize detection dataset

MaizeHorizon is a per-plant maize detection dataset recorded from a forward-facing RGB camera on a ground robot driving along crop rows toward the horizon. This viewpoint makes detection a far-field small-object problem: plants recede until they occupy a handful of pixels and merge into soil and canopy clutter.

The record accompanies the paper Far-Field Per-Plant Maize Detection: Network Input Size Does Not Lower the Floor, and is built to support its central measurement: a dissociation between the pixels the sensor resolves on a plant and the network input size.

Contents

• 5,979 RGB frames at 1920x1080 across 28 forward-motion clips, collected in two maize fields in Dan Phuong, Hanoi, Vietnam over three days.
• Hand-labeled test set: 3 held-out clips, 120 frames, 4,067 plant instances and 941 ignore regions, stratified by pixels-on-target into near (2,747), mid (1,238) and far (82) tiers.
• An explicit ignore class covering the unresolvable green band along the horizon, so that far-field evaluation neither penalizes correct detections nor rewards misses.
• Clip-disjoint splits for two training arms, preventing temporal leakage between near and far views of the same plant.
• Optically minted labels: 3,636 far-tier pseudo-label files and 10,355 near/far pairs of the same physical plant, obtained by detecting each plant when near and tracking it backward with optical flow.
• Metadata: a human-verified per-plant count ledger and the test-set box-area cumulative distribution.

Files

MaizeHorizon-images.tar (5.7 GB) holds the frames; MaizeHorizon-annotations.tar.gz (3.0 MB) holds labels, splits and metadata and is usable on its own. Both extract into a single data/ directory. Verify with SHA256SUMS.txt; see README.md for the full layout, label format and usage.

Code

Analysis and training code: github.com/manhhv87/MaizeHorizon (AGPL-3.0, inherited from Ultralytics YOLO). The data here are CC BY 4.0.

Scope

The data come from a single locality. Absolute recall, AP and counting figures characterize this domain rather than guaranteeing cross-domain transfer, and no claim of season or cultivar invariance is made.