I’m thinking about how to apply computer vision to basketball in order to quantify things that are not currently or see the game in a different way.

This post merely includes some early thoughts and miniscule prototypes.

I used Facebook Research’s DensePose. It was kind of a pain to install, so I wrote a little gist for it if you’re interested in tinkering yourself. It might not ultimately be the right tool though.

I didn’t pre-process the videos at all. The court and crowd definitely needs to be segmented out. Here’s what it looks like without that.

This isn’t NBA specific. Here I pulled a Euroleague clip. Also, I thought it might be interesting to turn the players into just a single-colored silhouette. Can that remove some bias in scouting? Can you guess who’s shooting in this clip?

This is very much a work in progress. Would love to hear any ideas or feedback!