AwnPhen Canonical YOLO11N Segmentation Model

This repository contains the canonical detection model used by AwnPhen and Awn Studio for automated wheat awn phenotyping from scanned spike images.

Intended use

The model provides instance-segmentation evidence for two classes:

awn
spikelet

Its masks are not the final phenotyping output. AwnPhen applies reconciliation, physical reconstruction with Unified Growth, representative-awn selection, calibration, and measurement after model inference.

Model

Architecture: YOLO11N segmentation
Input size used for the canonical pipeline: 640 px
Classes: awn, spikelet
Canonical tile stride: 320 px
Training epochs: 80
Training batch size: 20
Training seed: 20260906
Checkpoint filename: best.pt
SHA-256: a7a5cf23bf5d35266e4fa6b1dc0244ee802026a381548bcd202f04b3ebf42097

The training configuration is included as training_config.yaml.

Training and evaluation boundary

Model development used the project's training and validation partitions. The locked test set was not used for training or model selection.

This model is released as the canonical inference checkpoint for the AwnPhen pipeline. Model masks alone should not be interpreted as final awn-length measurements.

Expected input

Scanned wheat spike images compatible with the AwnPhen acquisition and calibration workflow. Physical length output requires valid image calibration.

Limitations

Performance can degrade under acquisition conditions that differ substantially from the development data, including unusual backgrounds, severe blur, extreme scale changes, or structures outside the target wheat-spike domain. Long thin awns can also be fragmented at the segmentation stage; the downstream Unified Growth reconstruction exists specifically to address this structural failure mode.

Use with AwnPhen

After the public Hugging Face repository ID is configured in AwnPhen:

awnphen demo
awnphen predict path/to/image.jpg
awnphen studio

For offline use, AwnPhen also accepts an explicit local checkpoint with --weights or the AWNPHEN_WEIGHTS environment variable.

License

This checkpoint was trained using Ultralytics YOLO. Ultralytics states that its YOLO trained models are distributed under AGPL-3.0 by default unless covered by an applicable commercial license. The public AwnPhen release should therefore be reviewed and licensed consistently before publication.

Downloads last month: 23