Inspiration

We aren't musicians. We can't dance. With AirTunes, we can try to do both! Superheroes are also pretty cool.

What it does

AirTunes recognizes 10 different popular dance moves (at any given moment) and generates a corresponding sound. The sounds can be looped and added at various times to create an original song with simple gestures. The user can choose to be one of four different superheroes (Hulk, Superman, Batman, Mr. Incredible) and record their piece with their own personal touch.

How we built it

In our first attempt, we used OpenCV to maps the arms and face of the user and measure the angles between the body parts to map to a dance move. Although successful with a few gestures, more complex gestures like the "shoot" were not ideal for this method. We ended up training a convolutional neural network in Tensorflow with 1000 samples of each gesture, which worked better. The model works with 98% accuracy on the test data set.

We designed the UI using the kivy library in python. There, we added record functionality, the ability to choose the music and the superhero overlay, which was done with the use of rlib and opencv to detect facial features and map a static image over these features.

Challenges we ran into

We came in with a completely different idea for the Hack for Resistance Route, and we spent the first day basically working on that until we realized that it was not interesting enough for us to sacrifice our cherished sleep. We abandoned the idea and started experimenting with LeapMotion, which was also unsuccessful because of its limited range. And so, the biggest challenge we faced was time.

It was also tricky to figure out the contour settings and get them 'just right'. To maintain a consistent environment, we even went down to CVS and bought a shower curtain for a plain white background. Afterward, we realized we could have just added a few sliders to adjust the settings based on whatever environment we were in.

Accomplishments that we're proud of

It was one of our first experiences training an ML model for image recognition and it's a lot more accurate than we had even expected.

What we learned

All four of us worked with unfamiliar technologies for the majority of the hack, so we each got to learn something new!

What's next for AirTunes

The biggest feature we see in the future for AirTunes is the ability to add your own gestures. We would also like to create a web app as opposed to a local application and add more customization.

Built With

  • open-cv
  • tensorflow
  • python
  • kivy
  • convolutional-neural-network
  • keras
Share this project:
×

Updates