maxdv/twostepper

TwoStepper

Hi! Learning a new dance move can be tough at any skill level. I developed TwoStepper to help dancers learn new moves. I always found that it was difficult when I was out at my favorite honkytonk to see a new move and remember how to do it later. I would always mess something up, or it wasn't quite right. TwoStepper exists so you can see the move, identify it by its name, and learn how to do it with a curated search in YouTube.

TwoStepper was built as a part of a three-week project for Insight Data Science in January 2020.

Visit TwoStepper at www.twostepper.xyz!

Working with the Data

Classifying dance moves has several challenges. Dancers can be viewed from any angle, any one move can take different amounts of time, and collapsing a time series into an input that is consistently recognizable is not an intuitive task.

Initially, I tried to collect a dataset of dancers dancing a select number of moves by recording the time that I saw it happen in existing tutorials on YouTube for three different moves:

• The Turn
• The Cuddle
• The Shadow

These YouTube queries resulted in lots of tutorials, and lots of instances within each video of the dancers moving around a studio from a variety of angles and proximity to the camera. Still, the process was labor-intensive, and required a lot of scrubbing back and forward within videos.

To augment the dataset, I decided to create a moving window around each known instance of the move happening in the video. One that started late, one that started early, and one that was a little long. With my clip database quadrupled in its number of elements, I then split each clip into a maximum of 30 evenly timed frames. For more details, look for the function videos_from_database in, 'classifier/frame_processer_functions.py'