Dataset · Computer vision · 2026

ChickenVerse

Name: ChickenVerse
Creator: Ahmad Amirivojdan
Published: 2026
License: https://creativecommons.org/licenses/by-nc-sa/4.0/

An open, large-scale multi-task dataset for chicken detection, instance segmentation, and behavior recognition.

Grid of overhead poultry-house frames with every bird outlined by a colored segmentation mask, grouped into train, validation and test farms.

6,539annotated images
153,764bird instances, box + mask
15,250single-bird video clips
15behavior classes
5poultry facilities

ChickenVerse is an openly released dataset for computer vision in broiler welfare monitoring. It combines imagery and video from five poultry facilities into two complementary sub-datasets, ChickenDet and ChickenAct, each shipped with recommended train/validation/test splits and evaluation protocols.

Most poultry vision datasets come from a single farm, so models trained on them rarely survive a new barn. ChickenVerse is split by facility on purpose, so the reported test scores measure how well a model transfers to a house it has never seen.

ChickenDet: detection and segmentation

ChickenDet is a COCO-format dataset in which every bird carries a mutually consistent bounding box and pixel-level mask. Frames were sampled from overhead recordings at five-minute intervals to maximize visual diversity, and masks were labeled in CVAT with SAM 2 assistance.

The training set spans multiple farms and camera geometries, validation comes from the JARTU research pens, and the test set is a fully held-out commercial broiler house (F4).

Split	Images	Instances	Mean per image
Train	5,222	116,329	22.3
Validation	1,067	24,815	23.3
Test	250	12,620	50.5
Total	6,539	153,764	23.5

ChickenAct: behavior recognition

ChickenAct is a video classification dataset of short single-bird clips covering 15 welfare-relevant behaviors, such as eating, drinking, walking, resting, preening and dust bathing. Clips were produced automatically by a detect-track-crop pipeline (a ChickenDet-finetuned RT-DETRv2 detector plus a Norfair tracker) and then labeled by hand in a lightweight annotation interface.

Sample clips per behavior class (first 12 seconds of the preview video).

Split	Clips	Classes
Train	11,278	15
Validation	1,829	15
Test	2,143	15
Total	15,250	15

Benchmarks

Both sub-datasets come with baselines. For behavior recognition, nine video models were compared on the held-out test split. V-JEPA 2 performed best, with 0.898 accuracy and 0.887 macro-F1.

Bar chart of test accuracy and macro-F1 for nine video models, led by V-JEPA 2 at 0.898 accuracy and 0.887 macro-F1. — ChickenAct test accuracy and macro-F1 by model.

Scatter plots of detection mAP50-95 against model size and inference time for YOLO variants and RT-DETR-L. — ChickenDet detection: accuracy versus model size and inference speed.

Scatter plots of instance segmentation accuracy against model size and inference speed. — ChickenDet instance segmentation: accuracy versus model size and inference speed.

License

The dataset is released under CC BY-NC-SA 4.0. You can share and adapt it for non-commercial purposes with credit, as long as derivatives use the same license.