I'm a big basketball fan (mostly college). Where are you sourcing the videos? Also, are you manually tagging each video with player names, teams, and keywords or is there some magic happening on the backend?
Just from a cursory glance at the results, it seems like they're simply providing a naive search on the play-by-play descriptions that the NBA provides (e.g. http://data.nba.net/data/10s/prod/v1/20161212/0021600366_pbp...). No computer vision behind-the-scenes magic, unfortunately.
> No computer vision behind-the-scenes magic, unfortunately.
Theoretically, it shouldn't be terribly technically difficult to attach a few sensors to each player to get a relatively hi-res 3D representation of each player's position throughout the game, should it?
Annotating the action accurately, not so easy, or might there be some trick?
This sort of happens already in NBA games. There's a system called SportVU that tracks XY positioning of players and XYZ positioning of the ball in real-time. Here's a decent article that describes it in more detail: http://www.fastcodesign.com/1670059/moneyball-20-how-missile... . Unfortunately, the data is proprietary.
my impression, from having talked with people who analyze the data from sportVU, is that the raw data is fairly useless and most of the useful data crunching happens overnight after each game.
with that said, bluetooth beacons, along with 3 bluetooth receivers for triangulation, could be used for inexpensive player tracking, but the data would be noisy and not as accurate (probably ~1 ft resolution or so). i've actually been wanting to put such a system together to determine how good it might be.