(9/14) Posenet adventures.

I had quite a fortuitous moment that my project was to focus on machine learning while our final projects were due. I have always been quite wary of using anything that does machine learning as I don’t think I am don’t have enough programming chops but ml5js laid that to rest. Picking it up was so easy and getting posenet to work was trivial. It gave me the confidence to approach these algorithms and start thinking on those lines.

Back to my concept.

I have always wanted to work with the body as an input (Since I don’t move so well in real life.) Until now, most of the work related to body, gait and posture have been with a kinect or a specialized camera that is currently beyond my scope of technical ability. But posenet makes body-tracking so trivial that I realised that I could work with it.

The posture is so iconic for a person that people are defined by it.

Screenshot 2018-11-07 08.50.58.png

And who hasn’t done this!

1920px-EB_Games_Expo_2015_-_Just_Dance_2016 (1).JPG

So i want to take the energy and goofy fun of dance games, karaoke and combine it with posture and movement.

My idea: People are given famous posture to imitate one after the other and they are supposed mimic it. The computer compares their posture to the original posture and scores them on the basis of it.


To get it working, I realised that I had to work with ratios and angle instead of absolute numbers. My explorations with the posenet library was to use it to calculate angles between various body parts but I was thrown off by the vector method in p5 which refused to work for me. So, here is a sketch of my failure: