Self-Supervised Representations for Tracking Data

Building tools that leverage tracking data to help accelerate video analysis workflows.
- Karun Singh (@karun1710)

Summary

Video analysis is a well-established and critical process in football, but meticulously studying match footage is incredibly time-consuming. The purpose of this project is to build collaborative tools that help analysts spend less time laboriously sifting through footage, and more time putting their deep domain expertise to use. We build 3 such tools that enable analysts to get instant responses to the following requests:

  1. Situation Search: "Find me clips of other instances where something similar happened."
  2. Auto-Tagging: "Tag all of Team X's counter-attacks this season."
  3. Team Fingerprints: "What are some things that Team X does unusually often?"

I recently had a chance to present this work at the Opta Pro Forum 2020 (thank you to all the organizers!) — you can find the presentation video, slides, a podcast interview, and details about a demo below. Please reach out if you have any thoughts or feedback!

Presentation Video

Slides

Presentation Slides (Google Drive)

Podcast

Thank you to Paul Carr and TruMedia for having me on their podcast, Expected Value! You can listen to the episode below.

Demo

If you're interested in trying out the tools yourself, please reach out to me on Twitter at @karun1710 or via email at karun.singh17@gmail.com!