> Also, the watch presumably can't tell how fast the ball is served, which depends on the racquet speed and whether you make good contact with the ball.
I think I disagree here. It should be possible to build a model that infers things like racquet speed and where the ball made contact with the racquet based on accelerometer data. At the end of the day you have an arm attached to a racket moving at some speed, that impacts a ball that is relatively not moving.
A straightforward model could calculate the lost momentum in the arm + racquet, assume the ball's mass, and work backwards to the ball's velocity from there. I'm assuming there is more ML magic behind Apple's implementation though.
To mix a metaphor, all good serves are the same; all bad serves are bad in their own way. A model might work well for people who generally serve well, or for an individual person who serves badly in a particular way. But how do you make a model that works for many different types of people, and many different types of serves?
It would be great to see if the data that the Apple Watch generates is similar to what speed guns show. That would make a great experiment for a high schooler!
I think I disagree here. It should be possible to build a model that infers things like racquet speed and where the ball made contact with the racquet based on accelerometer data. At the end of the day you have an arm attached to a racket moving at some speed, that impacts a ball that is relatively not moving.
A straightforward model could calculate the lost momentum in the arm + racquet, assume the ball's mass, and work backwards to the ball's velocity from there. I'm assuming there is more ML magic behind Apple's implementation though.