What's the structure of the data? about dagnet HOT 5 CLOSED

Hi,
thanks for your interest in our research!

Take a look at datasets/README.md: as you said, the first dimension is the number of sequences and the second is number of timesteps; the former dimensions contain a channel for each player (11: 5 attackers + 5 defenders + ball) and 4 channels to represent their position in the court. The 4th channel in the last dimension can always be dropped (if I recall correctly, when I downloaded the data for the first time they said it was a boolean flag for the player with ball possession, but I didn't found it correct), and since we're looking at players positions from a top-down view, we can also discard the z-channel keeping only xy.

Alex

from dagnet.

Thank you so much for your kind answer😀!

from dagnet.

Hello, I have one more question.

According to your paper, the NBA dataset is from “SportVU - STATS Perform”.
But I couldn't find more details there. May I know if it is the same as the data on this github link?

Thank you.

from dagnet.

Hi,
yes unfortunately SportVU turned the dataset private for commercial purposes a few years ago, hence we can only rely on mirrors or GitHub repos. Personally, I found the dataset here; I can't ensure the data from your link are exactly the same, but these versions as many others here on Github are all rip-offs of the original dataset, so they are surely coherent.

Alex

from dagnet.

Thank you again for your answer!

from dagnet.