The Big Broiler Dataset

In our earlier work, “Label Efficient Multi-Camera Broiler Localisation”, we used multiple cameras to detect and track broilers across a pen. While detection and tracking can tell us where birds are and how they move at a coarse level, that information alone does not give us the detailed behavioral signals we would need to detect disease in chickens.

To bridge this gap, we are introducing Big Broiler, a novel video dataset designed to train and evaluate algorithms for broiler behavior monitoring.

Why Action Segmentation?

If we want to monitor the health and welfare of a flock, we need to understand exactly what the chickens are doing and for how long. This is where action segmentation comes in. Instead of just asking “where is the chicken?” or “what is the chicken doing right now?”, action segmentation asks “what is the chicken doing, and exactly when did that behavior start and stop?”.

By classifying every single frame of a video, we can calculate the exact durations of specific behaviors—such as eating, drinking, resting, and preening. This allows us to establish a baseline of what is “normal” for a healthy flock. When a disease or environmental stressor affects the birds, their behavior often changes before visible clinical symptoms appear. By continuously tracking these behavior durations, we can detect deviations from the norm and flag potential problems early.

A Novel Task: Multi-Subject Action Segmentation

While action segmentation is well-studied for single human subjects (like analyzing cooking videos), it is rarely applied to multiple identical subjects simultaneously. The Big Broiler dataset introduces a novel, combined task: multi-subject tracking coupled with temporal action segmentation.

This means the system must first track individual broilers in a crowded, visually uniform environment, and then simultaneously segment the timeline of behaviors for each tracked bird. To our knowledge, no existing dataset provides this level of multi-object tracking combined with continuous, frame-wise behavior annotations.

Scale and Details of the Dataset

The Big Broiler dataset is massive in scale. It encompasses 102 videos recorded across three different experimental setups (varying pen sizes and camera configurations). In total, it provides:

• 898,607 labeled frames
• 3,659 unique broiler tracks
• 6.67 million frame-level activity annotations
• 28 unique behavioral and postural classes

The ethogram includes postural states (standing, sitting, lying), consumptive behaviors (eating, drinking), comfort behaviors (preening, dust bathing, stretching), and clinical signs (panting, sneezing) that are critical for welfare assessment.

Building the Dataset

Creating a dataset of this scale required precise temporal annotations of multiple subjects simultaneously. To achieve this, we developed a custom annotation tool tailored for multi-subject action segmentation in dense environments.

To ensure the quality of these labels, we conducted an Inter-Annotator Agreement (IAA) study. Three independent annotators labeled a holdout test set, and their agreement was evaluated to establish a reliable “ground truth” consensus.

This detailed labeling process allows us to train models that don’t just draw bounding boxes, but actually understand the temporal dynamics of each bird’s actions.

Models in Action

To see the potential of this dataset, here is an example of inference from the pipeline created by training on the Big Broiler dataset:

Get the Dataset

We are making the Big Broiler dataset available to the research community to help advance automated welfare monitoring in precision livestock farming.

You can find more information, including download links and the code repository, on our dedicated Big Broiler Dataset Page.