Williamsanderson/PoultryVision-Dataset

PoultryVision Unified Dataset

A large-scale, multi-modal poultry-farm dataset unifying six public sources for detection, classification, multi-camera tracking and behavior analysis of chickens (broilers, hens, cocks) and eggs.

This dataset was built to train Williamsanderson/PoultryVision, a YOLOv11m model that beats the fine-tuned YOLOv11x reported by Cardoen et al. (MVBroTrack paper, 2025) by +8.5 mAP@50-95.

---

📊 Dataset at a glance

Object detection (YOLO format)

Split	Images
Train	15 987
Val	3 706
Test	1 893
Total	21 586

Image classification

Split	Images
Train	1 832
Val	444
Test	263
Total	2 539

Videos & multi-camera

• 24 MP4 videos from 4 synchronized cameras (cam 9 / 10 / 11 / 12) across 6 samples
• Camera calibration files (intrinsics + extrinsics) for every camera
• Reprojection masks defining the ground-plane region of interest
• Tracking ground truth on the ground plane
• Pre-computed YOLO detections per frame for every multi-view sample

Classes (detection)

ID	Name	Description
0	chicken	All poultry: broilers, hens, cocks
1	egg	Chicken eggs

---

📚 Source datasets

#	Source	Type	Link / Reference
1	Dataset Chicken 1	Classification (images.cv)	images.cv
2	Dataset Chicken 2	Classification (images.cv)	images.cv
3	Dataset Chicken 3	Detection (COCO, Roboflow)	Roboflow Universe
4	Chickens-Eggs v1	Detection (YOLOv8, Roboflow)	Roboflow Universe
5	chicken eggs 2 v3	Detection + segmentation	Roboflow Universe
6	MVBroTrack	Multi-camera broiler tracking	Cardoen et al., Computers and Electronics in Agriculture, 2025

All six sources were standardized to a unified YOLO detection format (and/or ImageFolder classification format), deduplicated, and split into train/val/test.

---

🗂️ Folder structure

PoultryVision-Dataset/
├── data.yaml                      # Ultralytics-compatible dataset config
├── images/                        # Detection images
│   ├── train/   (15 987)
│   ├── val/     ( 3 706)
│   └── test/    ( 1 893)
├── labels/                        # YOLO .txt labels (one per image)
│   ├── train/
│   ├── val/
│   └── test/
├── classification/                # ImageFolder layout
│   ├── train/<class>/*.jpg
│   ├── val/<class>/*.jpg
│   └── test/<class>/*.jpg
├── videos/                        # 24 MP4 videos (4 cameras × 6 samples)
├── calibrations/                  # Camera calibration (intrinsics + extrinsics)
│   └── cam_<id>/
│       ├── intrinsics/{cameraMatrix.txt, distCoeffs.txt}
│       └── extrinsics/{rvec.txt, tvec.txt}
├── multi_view_detection/          # Pre-computed per-frame YOLO detections
├── reprojection_masks/            # Ground-plane ROI masks
└── tracking_gt/                   # Multi-camera tracking ground truth

---

🚀 Quick start

Download

pip install huggingface_hub
hf download --repo-type dataset Williamsanderson/PoultryVision-Dataset --local-dir PoultryVision-Dataset

Train a YOLO detector

from ultralytics import YOLO
model = YOLO("yolo11m.pt")
model.train(
    data="PoultryVision-Dataset/data.yaml",
    epochs=70,
    imgsz=640,
    batch=16,
    optimizer="AdamW",
    lr0=0.001,
)

Reference YOLO recipe that produced the published model:

model: yolo11m.pt
epochs: 70
imgsz: 640
optimizer: AdamW
lr0: 0.001
hsv_h: 0.015
hsv_s: 0.7
hsv_v: 0.4
mosaic: 1.0
mixup: 0.1
close_mosaic: 10
auto_augment: randaugment

Image classification

from torchvision.datasets import ImageFolder
from torchvision import transforms

tfm = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.ToTensor(),
])
train = ImageFolder("PoultryVision-Dataset/classification/train", transform=tfm)

Multi-camera tracking

Calibration files follow the MVBroTrack paper convention. Each camera folder contains cameraMatrix.txt, distCoeffs.txt, rvec.txt, tvec.txt. The repository ships a full multi-view pipeline (Algorithm 1 & 2 of the paper, Tracking-by-Curve-Matching) — see Williamsanderson/PoultryVision model repo.

---

🏆 Benchmark

Model trained on this dataset (YOLOv11m, 70 epochs, imgsz 640, AdamW):

Metric	Value
mAP@50-95	0.793
mAP@50	0.971
Precision	0.934
Recall	0.934

Compared to the MVBroTrack paper (Cardoen et al., 2025):

Model	mAP@50-95	Params
YOLOv11x fine-tuned (paper)	70.8 %	56.9 M
YOLOv11m fine-tuned (ours)	79.3 %	20.1 M

---

⚖️ License

This dataset is released under CC-BY-4.0.

• The unified packaging, splits and labels harmonization are © 2025 Williams Anderson, CC-BY-4.0.
• Individual source datasets retain their original licenses:
  • MVBroTrack (Cardoen et al., 2025) — see the original paper and its data statement
  • Roboflow Universe datasets — typically CC-BY-4.0 (check each source)
  • images.cv datasets — CC-BY-4.0 / public domain
• Please cite the original sources if you use the corresponding subsets.

---

📚 Citation

@misc{williamsanderson_poultryvision_dataset_2025,
  title  = {PoultryVision: A Unified Dataset for Poultry-Farm Computer Vision},
  author = {Williams Anderson},
  year   = {2025},
  howpublished = {\url{https://huggingface.co/datasets/Williamsanderson/PoultryVision-Dataset}},
}

@article{cardoen2025mvbrotrack,
  title   = {Multi-camera detection and tracking for individual broiler monitoring},
  author  = {Cardoen, J. and others},
  journal = {Computers and Electronics in Agriculture},
  year    = {2025}
}

---

🙏 Acknowledgements

• Cardoen et al. (MVBroTrack) for the multi-camera broiler data
• Roboflow and images.cv communities for the chicken / egg datasets
• Ultralytics for the YOLOv11 framework that produced the reference model