Open Access Dataset

Extreme Kepoints Dataset

Free Extreme Keypoints Dataset

Humans in the Loop is happy to share that we are publishing a new dataset which will help companies developing pose recognition AI to deal with the most extreme body positions. This Extreme Keypoints Dataset consists of 480 images collected from copyright-free sources where each image contains one person in an extreme or contorted pose. The person’s joints are annotated with 24 separate keypoints, following the Open Pose format. The annotation of this dataset was performed by 12 Humans in the Loop trainees in the Democratic Republic of Congo as part of their trainings.

This Extreme Keypoints dataset is dedicated to the public domain by Humans in the Loop under CC0 1.0 license.

Dataset size

The dataset includes 480 images with 11,520 keypoints in total.


There are 24 classes, each one for one keypoint on the body:

  • “Nose”, “Neck”, “RShoulder”,  “RElbow”, “RWrist”, “LShoulder”,  “LElbow”, “LWrist”, “MidHip”, “RHip”, “RKnee”, “RAnkle”,  “LHip”, “LKnee”, “LAnkle”, “REye”, “LEye”, “REar”, “LEar”, “LBigToe”, “LSmallToe”, “LHeel”, “RBigToe”, “RSmallToe”, “RHeel”.

Access the dataset by filling in the form below