Search Engine for Sports Scenes

Fast sports-scene search engine

Based on deep learning, our technology enables extremely fast searches for similar game scenarios in sports, with a search time below 1 second for about 1 million scenes. The search doesn’t require any manual annotation, only position data for the ball and players, which is already collected in many sports. This not only finds game situations very quickly, but should also automatically evaluate the execution of those situations. Using reinforcement learning, it should eventually be possible to generate unique solutions for real game situations and ultimately propose the most promising ones.

Technology innovation – performance comparison

Fast searches in tracking data as you know them from Internet search engines.

~1 million scenes	Fraunhofer technology	Naive search in tracking data	Alternative method based on tracking data
Search time	< 0.5 seconds	> 1 hour	> 30 seconds
Advantages	Fast search Independent of player role	Most accurate search type
Disadvantages	Known approximation error	Extremely long search time	Unknown approximation error

Scientific publications

Löffler, C., Reeb, L., Dzibela, D., Marzilger, R., Witt, N., Eskofier, B. M., & Mutschler, C. Deep Siamese Metric Learning: A Highly Scalable Approach to Searching Unordered Sets of Trajectories ACM Transactions on Intelligent Systems and Technology (TIST). Accepted, not yet published.

Reeb, L., Dzibela, D., Marzilger, R., & Witt, N. Effiziente Suche und Bewertung von Szenen in Spielsportarten. spinfortec 2020 digital, 20.

For which applications is the search engine suitable?

Game analysis

Finding and evaluating crucial or otherwise interesting scenes in recordings of sports matches is a basic work step in match analysis performed by clubs and associations (cf. DFB Academy: DFB Academy Analyses) and in media reporting. Currently, corresponding labels are entered manually during or after a game; alternatively, video footage is painstakingly searched after a game.

Using them in this project should enable the added value of more effective training management and talent scouting, such as through the automatic identification of weak points or the automatic compilation of a player’s well or poorly solved game scenes.

Media production & reporting

Finding crucial scenes in sports games or compiling tactical patterns in recordings can greatly simplify or even truly enrich media coverage.

Our portfolio & opportunities for collaboration

Fast search for your tracking data

If you want to integrate the fast search, you can license the technology.

We will be glad to help you integrate it into your system and adapt it as necessary.

Similar challenge for your data assets (big data)?

If you also manage large volumes of data and want to make them searchable quickly, contact us!

In an initial meeting, we will explore together whether customization makes sense for your application data.

Do you need a consultation or training on how to apply AI methods?

This is something else we can help you with: we offer both training on the practical use of AI (link to the ML seminar) and consulting on the application of AI in sports (link to sports analytics).

S3Engine