eye-pose-v0

YOLO11n-pose model fine-tuned on CUB-200-2011 for wildlife bird head localization. Used as Stage-1 localization in the image-scoring-model pipeline for eye-focus scoring.

Keypoints (6)

Index	Name
0	beak
1	left_eye
2	right_eye
3	head_top
4	left_shoulder
5	right_shoulder

flip_idx: [0, 2, 1, 3, 5, 4]

Training

• Base: YOLO11n-pose
• Dataset: CUB-200-2011 converted to YOLO pose format (~10k train / 1.7k val)
• Epochs: 100 (imgsz 640, batch 16)
• Final validation (epoch 100):
  • Pose mAP50: 0.994
  • Pose mAP50-95: 0.983
  • Box mAP50: 0.994

Usage

from ultralytics import YOLO

model = YOLO("hf://synthet/eye-pose-v0/eye_pose_v0.pt")
results = model.predict("bird.jpg", imgsz=640)

Or with the eye_quality package:

pip install -e "git+https://github.com/synthet/image-scoring-model.git"
eye-quality score bird.jpg --weights eye_pose_v0.pt

Download weights:

huggingface-cli download synthet/eye-pose-v0 eye_pose_v0.pt --local-dir models/

Limitations

• Single-class bird detection with head keypoints; not a general animal pose model.
• Trained on CUB-200 studio-style bird photos; performance on distant field wildlife may vary.
• Eye localization only — focus/sharpness scoring uses separate heuristics in the parent repo.