roboflow / sports

computer vision and sports

sports

notebooks | inference | autodistill | maestro

👋 hello

In sports, every centimeter and every second matter. That's why Roboflow decided to use sports as a testing ground to push our object detection, image segmentation, keypoint detection, and foundational models to their limits. This repository contains reusable tools that can be applied in sports and beyond.

🥵 challenges

Are you also a fan of computer vision and sports? We welcome contributions from anyone who shares our passion! Together, we can build powerful open-source tools for sports analytics. Here are the main challenges we're looking to tackle:

• Ball tracking: Tracking the ball is extremely difficult due to its small size and rapid movements, especially in high-resolution videos.
• Reading jersey numbers: Accurately reading player jersey numbers is often hampered by blurry videos, players turning away, or other objects obscuring the numbers.
• Player tracking: Maintaining consistent player identification throughout a game is a challenge due to frequent occlusions caused by other players or objects on the field.
• Player re-identification: Re-identifying players who have left and re-entered the frame is tricky, especially with moving cameras or when players are visually similar.
• Camera calibration: Accurately calibrating camera views is crucial for extracting advanced statistics like player speed and distance traveled. This is a complex task due to the dynamic nature of sports and varying camera angles.

💻 install

We don't have a Python package yet. Install from source in a Python>=3.8 environment.

pip install git+https://github.com/roboflow/sports.git

⚽ datasets

use case	dataset
soccer player detection
soccer ball detection
soccer pitch keypoint detection

Visit Roboflow Universe and explore other sport-related datasets.

🔥 demos

football-ai.mp4

🏆 contribution

We love your input! Let us know what else we should build!

sports's Issues

Include ball in radar view?

Great project! Thanks for sharing. I hope to test it out more extensively in the near future.

Is it possible to include the ball position in the radar view? That would give a bit more context to the player movements.

UnicodeDecodeError: 'charmap' codec can't decode byte 0x8f in position 3227: character maps to <undefined>

Tracking using new Meta SAM2

https://github.com/facebookresearch/segment-anything-2

Meta just open-sourced SAM2 which is supposed to be state of the art in tracking - and can track video in real time
Might this help with player & ball tracking?

License missing

Consider specifying a license for this repo.

How to run all three models at once

I have setup and installed all dependencies but when I pass this argument

#python examples/soccer/main.py --source_video_path examples/soccer/data/001.mp4 --target_video_path examples/soccer/data/001_output.mp4 --device cuda

I made very minor changes to make it run on gpu rest code is same
It runs only the first PLAYER_DETECTION model
Maybe my argument is wrong please tell me how to run the model so I get output as displayed

Sport AI

RuntimeError: PyTorch is not linked with support for mps devices

I clone the project, and trying to do what is mentioned in examples/sports. I first installed the requirements, run main.py for PITCH_DETECTION but got this error.

ModuleNotFoundError: No module named 'sports'

Hi @SkalskiP,

I was running the main.py inside the examples/soccer in my local system and getting an error like the following:

I moved the main.py to outside along with setup.py
Fortunately it worked. I was thinking if possible you could give it a look.

Thanks

Cannot find reusable code

Hi!

I was wondering where the bulk of your logic is in this project. You mention in the ReadMe that the code is migrating, what does that mean?

Can you please add support for basketball?

It would be very interesting to see a pipeline for basketball court and player tracking. Looking forward to it.

Code / Documentation

Where can i find the code or the documentation to achieve the same results as the demo video provided. I couldn't find the main file where all the classes and helper function were called.

Camera calibration

First of all, congratulations for sharing such nice project with the community 🎉

I see that in your blog posts and in the README, it is mentioned "camera calibration", but what is actually done is the homography estimation, which is different than camera calibration.

Camera calibration and homography estimation are related, but are distinct concepts.

Is it possible to make the code work with two video inputs?

Each video would represent one side of the field. The goal is to map all players on the field. TV broadcast videos often miss players who are off-screen, so with two cameras, I could cover parts of the field that are not shown.

Support for Table tennis

It would be good if there was some example support for table tennis. There is a model detection model here: https://github.com/sisrfeng/TTNet. If that can be used, maybe adding support to figure out when a point starts and when it ends would be helpful in analyzing matches