bird-detect-v0

bird-detect (v0, v1)

YOLO11n detect models for bird bounding boxes (single class bird). Companion to synthet/eye-pose-v0 for subject localization when eye keypoints are not required (species crops, gating, counting).

Used with the image-scoring-model eye-quality detect CLI.

Class

v1

Why. v0 learned from CUB's well-framed, centred birds. On long-lens field frames it missed most small or distant birds and had never seen a bird-free frame.

Training.

• Start: bird_detect_v0.pt, fine-tuned 30 epochs (imgsz 640, batch 16, lr0 0.002).
• Data: CUB-200-2011 boxes (10,018 train images) plus 803 teacher-labelled field frames listed 5× per epoch (29% of each epoch):
  • 455 bird boxes and 385 bird-free frames;
  • labelled by the upstream Apache-2.0 RTMDet-tiny COCO detector used as a teacher (full frame plus 2×2 tiles; only COCO bird ≥ 0.50 makes a box; anything ambiguous is excluded);
  • thresholds calibrated on an owner-labelled cohort whose folders are excluded from training.
• Validation (CUB val + teacher val, final epoch): mAP50 0.994, mAP50-95 0.878, precision 0.991, recall 0.982. Not comparable with v0's CUB-only validation numbers: the set now includes harder field frames.

Field evaluation (339-frame owner-labelled cohort from a wildlife library, never used in training; presence per frame, Wilson 95% intervals):

v0 training

• Base: YOLO11n (yolo11n.pt)
• Dataset: CUB-200-2011 boxes via data/wildlife_bird_det (~10k train / 1.7k val)
• Epochs: 100 (imgsz 640, batch 16)
• Final validation (epoch 100): Box mAP50 0.994, mAP50-95 0.892, precision 0.993, recall 0.997

Usage

from ultralytics import YOLO

model = YOLO("hf://synthet/bird-detect-v0/bird_detect_v1.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"
huggingface-cli download synthet/bird-detect-v0 bird_detect_v1.pt --local-dir models/
python -m eye_quality detect bird.jpg --weights models/bird_detect_v1.pt

Limitations

• Single-class bird boxes only; not a multi-species detector.
• v1's field data comes from one photographer's library (North American birds, long lens); validate on yours.
• The teacher excludes very small, silhouetted or occluded birds rather than labelling them, so those remain weak.
• CUB labels are typically one bird per image; crowded frames need care.
• CUB-200-2011 is a research dataset; check its terms for your use.