Dairy Goat Detection and Pose Estimation Dataset

This repository contains a unified dataset of annotated images of dairy goats recorded in an indoor barn environment. The dataset is designed to support computer vision research on automated detection of aggressive behavior in dairy goat herds.

The dataset combines two complementary annotation subsets, both encoded in COCO format:

Keypoint annotation subset: 1,107 PNG images with 11,687 annotations. Each annotation localizes up to 14 anatomical body-part keypoints per goat (nose, left eye, right eye, left hornbase, right hornbase, left eartip, right eartip, shoulder backline, tail base, tail tip, left carpus, right carpus, left tarsus, right tarsus). Keypoints were annotated using DeepLabCut and converted to COCO keypoint format.

Bounding box annotation subset: 606 PNG images with 10,933 annotations across three object classes (goat, head, body). Bounding boxes were annotated using CVAT and converted to COCO object detection format.

The images are frames extracted from video recordings captured by six overhead cameras installed at approximately 2.8 m above the floor of a dairy goat barn in Germany. Recordings were captured between April 2024 and August 2025 under natural and artificial lighting conditions. Any incidental human presence in the background has been fully anonymized via opaque masking.

The two image subsets are distinct. They do not share the same frames, although both are drawn from the same pool of barn video recordings. Not all goats visible in each image are annotated; annotators selected instances based on clarity and degree of occlusion. Similarly, not all 14 keypoints are labeled for every goat instance; the visibility flag in each keypoint triplet encodes which landmarks were annotated.

In addition to images and annotations, the repository includes:

Conversion scripts to reproduce the COCO JSON files from the original DeepLabCut CSV and CVAT XML annotation exports.

Marking scripts to visualize keypoint and bounding box annotations overlaid on images.

Integrity verification scripts that check COCO schema compliance, reference integrity, coordinate bounds, and other structural constraints.

Statistical validation scripts that analyze annotation distributions (keypoints per instance, annotations per image, bounding box sizes, aspect ratios, and overlap statistics), with accompanying reports and figures.

Dataset splitting scripts for creating train/validation/test subsets with proper annotation and image ID remapping.

Example machine learning scripts demonstrating training and evaluation of YOLOv8, RT-DETR, and SimpleBaseline models on the bounding box and keypoint tasks, respectively.

Real-time inference scripts for visualizing model predictions on streamed barn video.

Barn video recordings for use with the inference scripts.

Full documentation, including annotation guidelines, annotation definitions, and a description of the barn facility and camera installation.

A Datasheet for Datasets document (following Gebru et al., 2021).

All scripts are written in Python and include usage instructions in their headers. A requirements.txt file is provided for environment setup.

This work was funded by the Federal Office for Agriculture and Food of the Federal Republic of Germany as part of the VerZi (Automatische Verhaltensbewertung bei Milchziegen) project.