Dataset Card for Chicks4FreeID

The very first publicly available dataset for chicken re-identification.

1 Dataset Details

1.1 Dataset Description

The Chicks4FreeID dataset contains top-down view images of individually segmented and annotated chickens (with roosters and ducks also possibly present and labeled as such). 11 different coops with 54 individuals were visited for manual data collection. Each of the 677 images depicts at least one chicken. The identities of the 50 chickens, 2 roosters and 2 ducks were annotated for a total of 1270 animal instances. Annotation additionally contains visibility ratings of "best", "good", and "bad" for each animal instance. Besides chicken re-identification ,the curated dataset also support semantic and instance segmentation. Corresponding masks for these tasks are provided.

For a detailed description and documentation please read the paper and the supplementary material.

• Curated by: Daria Kern and Tobias Schiele
• Language(s) (NLP): English
• License: CC-BY-4.0

1.2 Dataset Sources

Code: GitHub

Paper: Link to OpenReview coming soon...

Supplementary material: Link to the supplementary material of the paper coming soon...

DOI: https://doi.org/10.57967/hf/2345

2 Uses

2.1 Direct Use

2.2 Out-of-Scope Use

Do not use for re-identification of roosters or ducks.

3 Dataset Structure

Modalities:

• 677 images
• 1270 preprocessed cut-out crops
• 1270 binary instance segmentation masks
• 677 color-Coded semantic segmentation masks (classes: chicken, rooster, duck, background)

Annotations:

• Animal category (chicken, rooster, duck)
• Identity (54 unique names)
• Coop (1-11)
• Visibility (best, good, bad)

4 Dataset Creation

4.1 Curation Rationale

The Chicks4FreeID dataset was created specifically for the task of chicken re-identification - i.e., recognizing the identity of an individual chicken in an image. There were two primary motivations for developing this dataset. First, there is a significant need for publicly available and well-annotated datasets in the field of animal re-identification. Second, there was a notable gap, as no such dataset existed for chickens prior to this effort.

However, the dataset is multipurpose and can also be used for semantic segmentation, instance segmentation, or even anomaly detection. It was structured, annotated, and prepared to support these additional tasks effectively.

4.2 Source Data

4.2.1 Data Collection and Processing

Data was collected manually using two models of cameras: the “Sony CyberShot DSC-RX100 VI”130 and the “Sony CyberShot DSC-RX100 I". The identities of the subjects were meticulously studied prior to photography, closely monitored throughout the image capture process, and ultimately assigned by a human annotator. No algorithms were used. During photography, the focus was always on a single chicken (the chickens were photographed sequentially, not randomly), while other individuals were able to enter the frame as well. The data collection took approximately one year. However, all images of a coop where always taken within a single day. In other words, all photos of an individual were taken on the same day.

4.2.2 Who are the source data producers?

Daria Kern collected the data.

4.2.3 Annotation process

We utilized Labelbox under a free educational license for manual data annotation.

4.2.4 Who are the annotators?

Daria Kern annotated the data.

4.2.5 Personal and Sensitive Information

The dataset does not contain any personal or sensitive information. It contains images of free-range chickens.

5 Bias, Risks, and Limitations

• Changes in appearance over time were not captured as all images of a given chicken were taken on the same day
• Despite the variability, chicken breeds included in the dataset are not exhaustive
• Be aware of class imbalances: the number of instances ranges from 4 to 27 in the "best" visibility subset

6 Citation

BibTeX:

@misc{kern2024Chicks4FreeID,
      title={Chicks4freeID}, 
      author={Daria Kern and Tobias Schiele and Ulrich Klauck and Winfred Ingabire},
      year={2024},
      doi={https://doi.org/10.57967/hf/2345},
      note={under review}
}

7 Dataset Card Contact

[email protected]