The very first publicly available chicken re-identification dataset is available on 🤗 Hugging Face: huggingface.co/datasets/dariakern/Chicks4FreeID

pip install datasets

Load the data:

from datasets import load_dataset
train_ds = load_dataset("dariakern/Chicks4FreeID", split="train")
train_ds[0]

Output:

{'crop': <PIL.PngImagePlugin.PngImageFile image mode=RGB size=2630x2630 at 0x7AA95E7D1720>,
'identity': 43}

To establish a baseline on the dataset, we explore 3 approaches

1. We evaluate the SotA model in animal re-identification: MegaDescriptor-L-384, a feature extractor, pre-trained on many species and identities.

   timm.create_model("hf-hub:BVRA/MegaDescriptor-L-384", pretrained=True)

2. We train MegaDescriptor-L-384's underlying architecture; a Swin-Transformer, in the same way it has been used to build the MegaDescriptor-L-384, but now on our own dataset.

   timm.create_model('swin_large_patch4_window12_384')

3. We train a Vision Transformer (ViT-B/16) as a fully supervised baseline, and focus on embeddings by replacing the classifier head with a linear layer.

   from torchvision.models import vit_b_16

Evaluation settings are based on:

• Linear: SimCLR
• k-NN: InstDist

Metrics are from torchmetrics

• mAP: MulticlassAveragePrecision(average="macro")
• top1: MulticlassAccuracy(top_k=1)
• top5: MulticlassAccuracy(top_k=5)

Below are the metrics for the test set. Standard deviations are based on 3 runs:

The most interesting observation in this table is that, even though the MegaDescriptor-L-384 feature extractor has never seen our dataset, its embeddings are still relatively helpful in identifiying the chickens, even when compared to the fully supervised approaches.

git clone https://github.com/DariaKern/Chicks4FreeID
cd Chicks4FreeID
pip install requirements.txt
python run_baseline.py

You can pass different options, depending on your hardware configuration

python run_baseline.py --devices=4 --batch-size-per-device=128 

For a full list of arguments type

python run_baseline.py --help

In a sepearte shell, open tensorboard to view progress and results

tensorboard --logdir baseline_logs

• [2024/05/30] DOI created: https://doi.org/10.57967/hf/2345
• [2024/05/23] the first version of the dataset was uploaded to Hugging Face. https://huggingface.co/datasets/dariakern/Chicks4FreeID

coming soon ...

@misc{kern2024Chicks4FreeID,
      title={Chicks4freeID: A Benchmark Dataset for Chicken Re-Identification}, 
      author={Daria Kern and Tobias Schiele and Ulrich Klauck and Winfred Ingabire},
      year={2024},
      doi={https://doi.org/10.57967/hf/2345},
      note={in preparation for NeurIPS 2024 Datasets and Benchmarks Track}
}