Video

does the model process a single frame at the time?

All the code for TeamClassifier is here: github.com/roboflow/sports

@@Roboflow Thank you very much!!!!

@@knobico1337 pleasure!

Important area of research! It is moving really fast over the past few years.

Thank you!

Stream processing is not the reason here. You’d need to train smaller architectures not YOLOv8x but YOLOv8s or m. But most importantly make embeddings calculation faster.

@@Roboflow thank you, i will try that my goal is to do live video feed processing. the concept is same where i need to capture keypoints and map it to a 2d map & mark trajectory but for a different project.

Hahah ;) I don’t think so

Do you mean „detect”?

Or each of the lights in the umbrella

Hi 👋🏻 looks to me like you didn’t executed all cells in order specifically missing one at the top that creates HOME variable.

Real football and basketball teams use data like this to for examples trade players looking for this that suit their play-style the best. Another usecase could be data overlay during game broadcasts or automated referees.

My guess is same reason why CharGPT responses are different every time you run it. Try adjusting temperature value.

Cool questions. Correct me if I’m wrong but optical flow won’t take me all the way. It only shows me that „something moved there” but does not tell me what it is. So if I want to know where are my reference points it won’t be able to do it. Because it does not know what points are interesting to me. Second of all pitch is not really static. You see pitch all the time but only some part of it. And you need a solution that will tell you what part of pitch is it. I didn’t really thought deeply about what model would I use to support 3D ball detection. But I recently was playing with 3D keypoint detection models that gave me 3D coordinates of human pose. I’d start there. But I’m open to other ideas.

Thanks for the swift answer. What I thought with optical flow is that given an initial state (1st frame) we can use anchor points known to be static in the world in order to estimate how the camera is moving. With that you can calculate how points in one frame relate to the next. But you are right, this approach is very fragile because it relies on having a set of anchor points which are always visible in the footage. For the 3D ball trajectory, because we know the physics of the ball movement, I’d probably start with trying a Kalman filter. One model for when the ball is in the air and another one for when on the ground. Use ball size in pixels as an estimate of how far from the camera it is. This is evidently very noisy but hopefully the dynamics of the kalman filter would smooth the trajectory

Hi you mean you can’t download video form Colab? Could you be a bit more specific?

You can find streams like that on YT for example. There are also street streams from NYC.

Great to hear that!

Awesome project! Would love to take a look at some visualizations from project like this! I’m currently working on tracks stitching. That would allow you to maintain same tracker ID.

@@Roboflow Waiting for a good outcome

Thanks a lot Ryan! 4 years in AI space feels like forever. I can’t even imagine what will happen over the next 4 years. It’s also super validating to hear you used similar strategies, and super interesting you’ve done this in completely different field. We are organizing community session next week. Is there any chance I could show some visualizations from your work on solar panels?

@@Roboflow Sure thing! Please send me an email and I can share images and more info

It is possible to run it on your own GPU (100% if you have Linux). I’m not sure about the windows part. I did not installed anything on windows in 15 years. :/

This is something that I clearly did not explained because you are not the first one to ask this question. So the problem is that every time different teams play. In one game red play against blue team. In different yellow against white team. There is no way you could annotate data in a way general enough you could apply it to any game in the future. Solution that I presented is general. And does not require you annotate all data with more info.

Ah that makes sense. Is soccer similar to other sports where the “away” team usually wears white/light jerseys and home wears dark/color jerseys? Wonder if home/away classes would be enough to generalize on

My pleasure!

Let me know if you like it and if you have some questions or cool ideas

Let me explain why I went with embeddings and not classification: problem with football is that every game is different. In one game you have white vs red. In other you have blue vs yellow. So I’d need to have separate classes for all of them. Embeddings is general solution. It doesn’t matter who plays. And it does not require annotation.

If you run it in Colab, I save in `/content` which is the default output directory

@@Roboflow So, all the output videos that you've shown in the tutorial will be saved in '/content' folder, right?

not sure what you mean, as I said multiple times in the video, already trained models are available for free, so you can skip the training part and go straight to inference

There are 100% easy optimizations that can bring us closer to… 15 FPS. Going faster than this can be challenging :)

It is really cool field to have career in!

You ask about specific Python libraries?

@@Roboflow yes. I'm new in it. So, i am taking to overview of the project. That's why ask for the technology used

I’ll try my best :)

Bro seriously of all the things you can learn or talk about, is how he speaks are you seriously Ok

@@judevector I'm not tryna offend him. Some people like me or just me find that so disturbing and can't focus....

It’s seriously not an issue. I didn’t even notice it. Don’t listen to this voice shamer.

I'll try to push out good content and not make that sound ;) it should be possible to do both at the same time haha or even better try to remove it in post production.

Let me know how you liked it

yup! all weights are saved in google colab after training; you can download them to your hard drive.

I think so! You can certainly try it and see how it goes for you! Google Colab does not require any installation.

yup! that would be a natural extension of that project, we know the accurate ball path so it should not be overly complicated

population of subscribers from Europe is increasing haha

@@SkalskiP You also got subscribers from China haha. In China, We call this sport 足球 which literally means football. 足 shares the same meaning with foot and 球 means ball.

the most obvious choice is traffic analysis in large-area stores; people want to know how customers move, where do they stop

@@SkalskiP thank you sir

That was a risky move. I bet it can work both ways haha

I see you decided not to skip model training section ;)

Haha! I decided that important things need to be explained right at the beginning.

Awesome to hear that! You came back after almost 2 years for part 2.