camarcano.bsky.social
5x @FSWA finalist, never a winner. Words: @baseballpro. Prev. @PitcherList.
Looking for speX? Here:
https://www.basebotics.com
Looking for Baseball Computer Vision tools? BaseballCV is here!
https://github.com/dylandru/BaseballCV
241 posts
2,577 followers
683 following
Prolific Poster
Active Commenter
comment in response to
post
Yep! That's something I'm testing, thanks for the feedback!
comment in response to
post
(4/4) We don’t currently have plans to generate more data—full-season tracking requires compute we don’t have. But you can generate it yourself using our BaseballTools class from BaseballCV:
github.com/dylandru/Bas...
comment in response to
post
(3/4) This is our first dataset that isn’t BaseballCV-based, so we’re still working on adding support in BaseballCV’s LoadTools. Expect direct access through the library soon!
comment in response to
post
(2/4) All data was generated by @dyland22.bsky.social and me using the GloveTracker from BaseballCV. While some plays may be missing due to processing, there’s more than enough here to start pulling out signal. Big thanks to @tangotiger.com for helping validate the early results against Hawkeye.
comment in response to
post
The famous "Judicial Effect".
comment in response to
post
That’s embarrassing.
comment in response to
post
Valid! Thanks!
comment in response to
post
Thanks! And yes, I see where are you coming from (we even got some encouraging comps with HawkEye data, thanks to @tangotiger.com ); definitely getting very granular would yield better results.
What surprised us, in a way, was that a monocular solution showed interesting possibilities.
comment in response to
post
I appreciate it, thanks!!!
comment in response to
post
That’s a great call, of course.
One of our current goals is to open source a huge amount of glove tracking data so more analysts can come up with analyses of their own and better ways of tackling the subject.
It takes a lot of time and processing but we are already on it!
comment in response to
post
This is exactly one of the multiple complicated instances that need to be tackled!
comment in response to
post
I appreciate the interest, Eno!
comment in response to
post
Thanks, Josh!
comment in response to
post
🙏🙏🙏
comment in response to
post
We are using especifically trained computer vision models to extract the data for the glove movement from the videos and then we process it all with a bunch of different algorithms to find the target. It is far from perfect at this moment but it's trending nicely, to be honest.
comment in response to
post
I think Command+ is from STATS, if I recall right. But you are right, it's pretty hard to get right. The conundrum to me is, it is one thing to get the measurement (what we are trying to achieve here) but then there is another thing in finding what to do with that data; I think that's the issue.
comment in response to
post
that it can be done, needing only a reasonable quality video.
That's what our CommandAnalyzer class, part of the Baseball Tools in BaseballCV aims to achieve, after these proof of concepts.
It'll be available soon from github.com/dylandru/Bas...
comment in response to
post
but the timing varies greatly; that's an issue. Also, some barely show intent location and then "hide" the glove, so these two things are opposing in how to handle them and mess up with each other, for example. There are a lot of things to take care of but the point is (4/)
comment in response to
post
Trying to catch, pun intended, the Target position of the glove, just by using the full sequence's glove position tracking, is not trivial as there are many implicit variables working at the same time.
A lot of catchers put the glove down after showing the intended location (3/)
comment in response to
post
After Dylan Drummey and I got a fully working catcher's glove tracking tool, we wanted to find useful applications for it and this is one that popped up immediately: Can "intend/actual" tracking be done by means of using only that glove data?
Well, it looks like it does. (2/)