Microsoft’s COCO (Common Objects in Context) dataset contains 220,000 labeled images, originally sourced from Flickr, and then tagged and annotated by workers on Amazon Mechanical Turk.

The goal of the project, which is used as training material for countless machine learning systems, is to enable state-of-the-art object recognition by “gathering images of complex everyday scenes containing common objects in their natural context.”

One such system is OpenPose, which is able to estimate human poses from still images.

Here I attempt to match my pose with a random selection of the original COCO images on which the system was trained. A python script chooses a random COCO image and subject for me to imitate, and I try to match its pose, using OpenPose to track myself in real time. When I succeed (within a threshold) a photo is automatically taken and then superimposed on the original.