The Big Broiler Dataset

The Big Broiler dataset provides extensive annotations for multi-subject action segmentation in broiler chickens. It is designed to facilitate automated behavior monitoring: by tracking the exact duration of activities such as eating, drinking, resting, and preening, researchers can establish a baseline of normal flock behavior and detect deviations that may signal health or welfare problems before clinical symptoms appear.

For background on why tracking alone is not enough for welfare monitoring, what multi-subject action segmentation means, and how we envision using behavior durations to spot early signs of problems, see our introductory blog post on the Big Broiler dataset.

Downloads & Resources

• Labelled data (~30 GB): Download (coming soon)
• Unlabelled data (~100 GB): Download (coming soon)
• GitHub Repository

Experimental Setups

Data were collected from three distinct experimental setups: a medium-scale setup (9x4 meter pen housing ~140 broilers; setup 1), a small-scale angled-view setup (pens with approximately 10 broilers each; setup 2), and a small-scale overhead setup (cameras mounted directly above; setup 3). Broilers were recorded from 10 to 40 days of age at 1920x1080 resolution and 25 fps.

Dataset Summary

The dataset comprises 102 videos with annotations for 27 behaviors plus an Out_of_sight label when birds are partially or fully obstructed. In total: 898,607 labelled frames, 3,659 broiler tracks, and 6,670,416 frame-level bounding boxes with activity labels.

Setup 2 is split into automated tracks (SORT + human filtering), a 62-track cross-validation set (manual tracks, tracker tuning), and a 99-track holdout test set (manual tracks, three concurrent annotators for inter-annotator agreement). Average track duration is average frames per track divided by 25 fps.