nrehiew

✌️

[nrehiew @nrehiew]

Ahead of the #WorldCup2022 finals, here is a look at what has been an incredibly entertaining World Cup

A Data Viz Thread 👇:

First, inspired by @petermckeever and @johnspacemuller, I start by looking at where each team enjoyed the most possession

[nrehiew @nrehiew]

@drscotthawley thanks for the reply! so its essentially at any single time t, it is impossible for any point to share the same position? Otherwise, it wont be invertible right?

[nrehiew @nrehiew]

@drscotthawley Thanks for the tutorial! Quick question on the crossing of the trajectories. it looks like some of the trajectories here also cross?

Maybe a clearer eg is with the target set to "2 Gaussians". Could this be because the model was not trained well?

[nrehiew @nrehiew]

@pranav_m28 thanks so much for the kind words ❤️

[nrehiew @nrehiew]

Just posted a thread showcasing the robustness of Eagle something I’m really proud of

[nrehiew @nrehiew]

One thing that Eagle 🦅 is strong at is just how robust it is to various settings/camera angles

The original demo video showed some variety so here's even more non-cherry-picked challenging examples...

https://t.co/vqLzz4KhmT

[nrehiew @nrehiew]

Excited to finally release Eagle 🦅

Eagle is a collection of models and tooling designed to easily extract tracking data from broadcast feeds which I hope can make tracking data and analysis more widely accessible to everyone

🧵

[nrehiew @nrehiew]

This is the first version of Eagle and I believe that a lot of work can be done in this area!

Any form of feedback/contribution is greatly appreciated ❤️

https://t.co/lJ4gPXh5nh

You can try it out immediately on Colab too! https://t.co/9UUI3H0mPe

[nrehiew @nrehiew]

All code can be found here!

https://t.co/lJ4gPXh5nh

It is also important to note that Eagle wasn't designed for visualisations but for analysis use cases which is why the coordinates are not smoothed out. By default, Eagle also only uses 24 FPS

[nrehiew @nrehiew]

Finally, my favourite and probably most challenging example. Sometimes, replays use this vertical moving camera.

Eagle works great too!

[nrehiew @nrehiew]

For transparency, I'll also add this example where Eagle is able to detect players in a crowded penalty box but gets confused with detecting the ball (especially since jersey colors are similar to the ball as well)

[nrehiew @nrehiew]

Now switching to something completely different, a tactical camera which is more suited for analysis use cases

[nrehiew @nrehiew]

For long shots, this behind-the-goal camera is often used

Eagle gets a bit confused with the ball but player coordinates are perfect!

[nrehiew @nrehiew]

One thing that Eagle 🦅 is strong at is just how robust it is to various settings/camera angles

The original demo video showed some variety so here's even more non-cherry-picked challenging examples...

https://t.co/vqLzz4KhmT

[nrehiew @nrehiew]

Excited to finally release Eagle 🦅

Eagle is a collection of models and tooling designed to easily extract tracking data from broadcast feeds which I hope can make tracking data and analysis more widely accessible to everyone

🧵

[nrehiew @nrehiew]

Another common camera angle used mainly for replays is a higher view. Eagle works perfectly fine too

[nrehiew @nrehiew]

Starting off with a classic sideline camera angle that follows the movement of play

Eagle is able to keep up with the movement accurately here with events happening across 2/3 of the field

[nrehiew @nrehiew]

@owen_seymour Hahah thanks owen for the kind words

[nrehiew @nrehiew]

@playingsquirrel thanks for the kind words! pretty surreal since I love your fm content :)

[nrehiew @nrehiew]

Excited to finally release Eagle 🦅

Eagle is a collection of models and tooling designed to easily extract tracking data from broadcast feeds which I hope can make tracking data and analysis more widely accessible to everyone

🧵

[nrehiew @nrehiew]

@omareltouny7820 Yea I tried a Resnet into transformer jersey number captioning thing but it was really bad at knowing when a jersey number is not visible so even when nothing was visible it would output

[nrehiew @nrehiew]

@LazioB I just did it with cv2/matplotlib haha