Chicks4FreeID: A Benchmark Dataset for Chicken Re-Identification

Daria Kern, Tobias Schiele, Ulrich Klauck, Winfred Ingabire

Keywords: animal re-identification, chicken re-identification, chicken dataset, re-identification dataset, benchmark dataset, closed-set re-identification, computer vision, chicken instance segmentation, poultry, livestock

TL;DR: The first publicly available dataset for chicken re-identification

Abstract: To address the need for well-annotated datasets in the field of animal re-identification, and particularly to close the existing gap for chickens, we introduce the Chicks4FreeID dataset. This dataset is the first publicly available re-identification resource dedicated to the most farmed animal in the world. It includes top-down view images of individually segmented and annotated chickens, along with preprocessed cut-out crops of the instances. The dataset comprises 1215 annotations of 50 unique chicken individuals, as well as a total of 55 annotations of 2 roosters and 2 ducks. In addition to re-identification, the dataset supports semantic and instance segmentation tasks by providing corresponding masks. Curation and annotation were performed manually, ensuring high-quality, nearly pixel-perfect masks and accurate ground truth assignment of the individuals using expert knowledge. Additionally, we provide context by offering a comprehensive overview of existing datasets for animal re-identification. To facilitate comparability, we establish a baseline for the re-identification task testing different approaches. Performance is evaluated based on mAP, Top-1, and Top-5 accuracy metrics. Both the data and code are publicly shared under a CC BY 4.0 license, promoting accessibility and further research. The dataset can be accessed at https://huggingface.co/datasets/dariakern/Chicks4FreeID and the code at https://github.com/DariaKern/Chicks4FreeID.

Flagged For Ethics Review: true

Submission Number: 1253