8-Calves Dataset

A benchmark dataset for occlusion-rich object detection, identity classification, and multi-object tracking. Features 8 Holstein Friesian calves with unique coat patterns in a 1-hour video with temporal annotations.

---

Overview

This dataset provides:

• 🕒 1-hour video (67,760 frames @20 fps, 600x800 resolution)
• 🎯 537,908 verified bounding boxes with calf identities (1-8)
• 🖼️ 900 hand-labeled static frames for detection tasks
• Designed to evaluate robustness in occlusion handling, identity preservation, and temporal consistency.

*Example frame with bounding boxes (green) and calf identities. Challenges include occlusion, motion blur, and pose variation.*

---

Key Features

• Temporal Richness: 1-hour continuous recording (vs. 10-minute benchmarks like 3D-POP)
• High-Quality Labels:
  • Generated via ByteTrack + YOLOv8m pipeline with manual correction
  • <0.56% annotation error rate
• Unique Challenges: Motion blur, pose variation, and frequent occlusions
• Efficiency Testing: Compare lightweight (e.g., YOLOv9t) vs. large models (e.g., ConvNextV2)

---

Dataset Structure

hand_labelled_frames/ # 900 manually annotated frames and labels in YOLO format, class=0 for cows

pmfeed_4_3_16.avi # 1-hour video (4th March 2016)

pmfeed_4_3_16_bboxes_and_labels.pkl # Temporal annotations

Annotation Details

PKL File Columns:

Column	Description
class	Always 0 (cow detection)
x, y, w, h	YOLO-format bounding boxes
conf	Ignore (detections manually verified)
tracklet_id	Calf identity (1-8)
frame_id	Temporal index matching video

Load annotations:

import pandas as pd
df = pd.read_pickle("pmfeed_4_3_16_bboxes_and_labels.pkl")

Usage

Dataset Download:

Step 1: install git-lfs: git lfs install

Step 2: git clone [email protected]:datasets/tonyFang04/8-calves

Step 3: install conda and pip environments:

conda create --name new_env --file conda_requirements.txt
pip install -r pip_requirements.txt

Object Detection

• Training/Validation: Use the first 600 frames from hand_labelled_frames/ (chronological split).
• Testing: Evaluate on the full video (pmfeed_4_3_16.avi) using the provided PKL annotations.
• ⚠️ Avoid Data Leakage: Do not use all 900 frames for training - they are temporally linked to the test video.

Recommended Split:

Split	Frames	Purpose
Training	500	Model training
Validation	100	Hyperparameter tuning
Test	67,760	Final evaluation

Benchmarking YOLO Models:

Step 1: cd 8-calves/object_detector_benchmark. Run ./create_yolo_dataset.sh and create_yolo_testset.py. This creates a YOLO dataset with the 500/100/67760 train/val/test split recommended above.

Step 2: find the Albumentations class in the data/augment.py file in ultralytics source code. And replace the default transforms to:

# Transforms
T = [
    A.RandomRotate90(p=1.0),
    A.HorizontalFlip(p=0.5),
    A.RandomBrightnessContrast(p=0.4),
    A.ElasticTransform(
        alpha=100.0, 
        sigma=5.0, 
        p=0.5
    ),
]

Step 3: run the yolo detectors following the following commands:

cd yolo_benchmark

Model_Name=yolov9t

yolo cfg=experiment.yaml model=$Model_Name.yaml name=$Model_Name

Benchmark Transformer Based Models:

Step 1: run the following commands to load the data into yolo format, then into coco, then into arrow:

cd 8-calves/object_detector_benchmark
./create_yolo_dataset.sh
python create_yolo_testset.py
python yolo_to_coco.py
python data_wrangling.py

Step 2: run the following commands to train:

cd transformer_benchmark
python train.py --config Configs/conditional_detr.yaml

Temporal Classification

• Use tracklet_id (1-8) from the PKL file as labels.
• Temporal Split: 30% train / 30% val / 40% test (chronological order).

Benchmark vision models for temporal classification:

Step 1: cropping the bounding boxes from pmfeed_4_3_16.mp4 using the correct labels in pmfeed_4_3_16_bboxes_and_labels.pkl. Then convert the folder of images cropped from pmfeed_4_3_16.mp4 into lmdb dataset for fast loading:

cd identification_benchmark
python crop_pmfeed_4_3_16.py
python construct_lmdb.py

Step 2: get embeddings from vision model:

cd big_model_inference

Use inference_resnet.py to get embeddings from resnet and inference_transformers.py to get embeddings from transformer weights available on Huggingface:

python inference_resnet.py --resnet_type resnet18
python inference_transformers.py --model_name facebook/convnextv2-nano-1k-224

Step 3: use the embeddings and labels obtained from step 2 to conduct knn evaluation and linear classification:

cd ../classification
python train.py
python knn_evaluation.py

Key Results

Object Detection (YOLO Models)

Model	Parameters (M)	mAP50:95 (%)	Inference Speed (ms/sample)
YOLOv9c	25.6	68.4	2.8
YOLOv8x	68.2	68.2	4.4
YOLOv10n	2.8	64.6	0.7

---

Identity Classification (Top Models)

Model	Accuracy (%)	KNN Top-1 (%)	Parameters (M)
ConvNextV2-Nano	73.1	50.8	15.6
Swin-Tiny	68.7	43.9	28.3
ResNet50	63.7	38.3	25.6

---

Notes:

• mAP50:95: Mean Average Precision at IoU thresholds 0.5–0.95.
• KNN Top-1: Nearest-neighbor accuracy using embeddings.
• Full results and methodology: arXiv paper.

License

This dataset is released under CC-BY 4.0.
Modifications/redistribution must include attribution.

Citation

@article{fang20248calves,
  title={8-Calves: A Benchmark for Object Detection and Identity Classification in Occlusion-Rich Environments},
  author={Fang, Xuyang and Hannuna, Sion and Campbell, Neill},
  journal={arXiv preprint arXiv:2503.13777},
  year={2024}
}

Contact

Dataset Maintainer:
Xuyang Fang
Email: [email protected]