docs/en/datasets/pose/coco.md
The COCO-Pose dataset adapts COCO (Common Objects in Context) for pose estimation: 58,945 images from COCO Keypoints 2017, annotated with 156,165 people using a 17-keypoint schema. It is the standard set for training and benchmarking keypoint models such as Ultralytics YOLO26, and the 8-image COCO8-Pose subset mirrors its format for quick sanity checks.
{% include "macros/yolo-pose-perf.md" %}
(x, y, visibility) triplets.coco-pose.yaml header lists 20.1 GB (train2017.zip + val2017.zip only), but the download script also fetches the 7 GB test2017.zip unconditionally, even though that archive is needed only for the optional test-dev2017 submission split.For training and validation, COCO-Pose includes only COCO 2017 images with keypoint-annotated people, so its labeled splits are smaller than full COCO's. Its YAML defines three subsets:
Training at this scale is where Ultralytics Platform helps most — it manages the compute so you can launch and monitor runs without provisioning your own GPUs.
The COCO-Pose dataset is specifically used for training and evaluating deep learning models on keypoint detection and pose estimation. The dataset's large number of annotated images and standardized evaluation metrics make it an essential resource for computer vision researchers and practitioners working on human pose.
A YAML (Yet Another Markup Language) file is used to define the dataset configuration. It contains information about the dataset's paths, classes, and other relevant information. In the case of the COCO-Pose dataset, the coco-pose.yaml file is maintained at https://github.com/ultralytics/ultralytics/blob/main/ultralytics/cfg/datasets/coco-pose.yaml.
!!! example "ultralytics/cfg/datasets/coco-pose.yaml"
```yaml
--8<-- "ultralytics/cfg/datasets/coco-pose.yaml"
```
To train a YOLO26n-pose model on the COCO-Pose dataset for 100 epochs with an image size of 640, you can use the following code snippets. For a comprehensive list of available arguments, refer to the model Training page.
!!! example "Train Example"
=== "Python"
```python
from ultralytics import YOLO
# Load a model
model = YOLO("yolo26n-pose.pt") # load a pretrained model (recommended for training)
# Train the model
results = model.train(data="coco-pose.yaml", epochs=100, imgsz=640)
```
=== "CLI"
```bash
# Start training from a pretrained *.pt model
yolo pose train data=coco-pose.yaml model=yolo26n-pose.pt epochs=100 imgsz=640
```
The COCO-Pose dataset contains a diverse set of images with human figures annotated with keypoints. Here are some examples of images from the dataset, along with their corresponding annotations:
The example showcases the variety and complexity of the images in the COCO-Pose dataset and the benefits of using mosaicing during the training process.
If you use the COCO-Pose dataset in your research or development work, please cite the following paper:
!!! quote ""
=== "BibTeX"
```bibtex
@misc{lin2015microsoft,
title={Microsoft COCO: Common Objects in Context},
author={Tsung-Yi Lin and Michael Maire and Serge Belongie and Lubomir Bourdev and Ross Girshick and James Hays and Pietro Perona and Deva Ramanan and C. Lawrence Zitnick and Piotr Dollár},
year={2015},
eprint={1405.0312},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
```
We would like to acknowledge the COCO Consortium for creating and maintaining this valuable resource for the computer vision community. For more information about the COCO-Pose dataset and its creators, visit the COCO dataset website.
COCO-Pose supplies the COCO Keypoints 2017 images and annotations converted to YOLO keypoint format, using a 17-keypoint schema across 58,945 images. Point any Ultralytics YOLO pose model at it with data=coco-pose.yaml, and the Training page documents every argument you can tune from there.
Load yolo26n-pose.pt and call model.train(data="coco-pose.yaml", epochs=100, imgsz=640) — see the Train Example above for the full Python and CLI snippets, and the training page for a comprehensive list of arguments.
The COCO-Pose dataset provides several standardized evaluation metrics for pose estimation tasks, similar to the original COCO dataset. Key metrics include the Object Keypoint Similarity (OKS), which evaluates the accuracy of predicted keypoints against ground truth annotations. These metrics allow for thorough performance comparisons between different models. For instance, the COCO-Pose pretrained models such as YOLO26n-pose, YOLO26s-pose, and others have specific performance metrics listed in the documentation, like mAP<sup>pose</sup>50-95 and mAP<sup>pose</sup>50.
COCO-Pose ships two labeled splits: 56,599 train2017 images and 2,346 val2017 images. A third split, test-dev2017 (20,288 of the full 40,670 test2017 images), keeps its ground truth private; the dataset YAML links it to the COCO test-dev keypoints evaluation server. See the Dataset Structure section, or the coco-pose.yaml file on GitHub for the exact split paths.
COCO-Pose uses 17 human keypoint types and inherits COCO's standardized metrics, including Object Keypoint Similarity (OKS), for comparing models. That combination suits human pose applications such as sports analytics, healthcare, and human-computer interaction. Pretrained YOLO26-pose weights are listed under COCO-Pose Pretrained Models.
For more on keypoint models, see the Pose Estimation task docs.