sheep-pose-yolo26n

SamSeesSheep — sheep-head keypoint detector (YOLO26n-pose)

A lightweight (2.5 M parameter, ~6 MB) YOLO-pose model that places five anatomical keypoints on every sheep head it detects in ambient pasture video: nose, left/right ear base, left/right ear tip. From those keypoints the project pipeline derives a per-animal ear-angle signal.

What this is — and is not. This is a measurement instrument: it localizes ear keypoints with quantified stability. It is not a pain detector and not a welfare scorer. Ear-angle thresholds drawn from the literature are reference bands, not diagnostic cutoffs. See VALIDATION.md for the full anti-overclaim contract.

Keypoint schema

kpt_shape: [5, 3] — each keypoint is (x, y, visibility). Left/right are from the camera's perspective.

Horizontal-flip augmentation uses flip_idx: [0, 2, 1, 4, 3] (swaps left/right).

Model versions

Each version is the same architecture and recipe (YOLO26n-pose, 100 epochs, batch 8, imgsz 640); the only variable is labeled-data scale. The progression is published as a training-scale study, not as monotonic improvement past v0.4.

Use v0.7 (latest, most data) unless you are reproducing a specific paper figure. On a second held-out clip (HO-2, morning light), v0.7 reaches σ_avg = 2.84° — the lower number reflects an easier clip (larger apparent head, higher contrast), not a model improvement. σ here is the residual standard deviation of ear angle on a near-stationary sheep, i.e. the measurement noise floor; the difference between adjacent versions is within bootstrap CIs.

Stock yolo26n.pt (COCO-pretrained) produces zero keypoints on these clips — the keypoint head must be trained against the target animals.

Usage

from huggingface_hub import hf_hub_download
from ultralytics import YOLO

path = hf_hub_download("antking1/sheep-pose-yolo26n",
                       "sheep-pose-v0.7-yolo26n.pt")
model = YOLO(path)

results = model("sheep.jpg")           # detect heads + 5 keypoints each
kpts = results[0].keypoints.xy         # (n_sheep, 5, 2) pixel coords

Scope and limitations

Trained on a single Katahdin hair-sheep flock (5 ewes) on one homestead in Middletown, Delaware, USA, with one smartphone camera, reviewed by one non-veterinarian operator. No cross-breed, cross-flock, or cross-geography generalization is claimed. Known failure modes: dark fleece / low contrast, heavy motion blur, profile views with ear occlusion, crowded frames, wool-covered ears. See the repository for the full benchmark protocol and the verify_paper_claims.py harness (59/59 checks).

Links

• Code + benchmarks: https://github.com/antonemking/SamSeesSheep
• Paper: SamSeesSheep: A Measurement Pipeline for Sheep Ear-Angle Extraction from Ambient Pasture Video Using Foundation-Model Annotation (see repository)
• Prediction caches (to regenerate every benchmark JSON): GitHub release asset

License

Released under AGPL-3.0 to match Ultralytics YOLO, with which these weights were trained. (The author's own pipeline code in the GitHub repository is MIT; the trained weights inherit Ultralytics' AGPL-3.0.)

Citation

@misc{king2026samseessheep,
  author       = {Antone King},
  title        = {SamSeesSheep: A Measurement Pipeline for Sheep Ear-Angle
                  Extraction from Ambient Pasture Video Using Foundation-Model
                  Annotation},
  year         = {2026},
  howpublished = {\url{https://github.com/antonemking/SamSeesSheep}},
}