The authors introduce the concept of Game State Reconstruction (GSR), a novel computer vision task that aims to automatically track and identify all athletes on a 2D minimap of the sports pitch, using single-camera video footage. To support research on this task, they release the SoccerNet-GSR dataset, which contains 200 30-second fully annotated video clips with over 9.37 million line points for pitch localization and camera calibration, as well as over 2.36 million athlete positions on the pitch with unique identification information including their role, team, and jersey number.
The authors also introduce GS-HOTA, a new evaluation metric to benchmark GSR methods, which accounts for both the localization and identification accuracy of the predicted athletes. Finally, they propose GSR-Baseline, the first end-to-end and open-source pipeline for game state reconstruction, built upon state-of-the-art tracking, re-identification, team affiliation, jersey number recognition, pitch localization, and camera calibration methods.
The analysis of the results underscores the complexity of the Game State Reconstruction task and highlights the importance of introducing this new benchmark. This initiative establishes an ideal platform for future research in the field, aiming to democratize access to this valuable game state data for all leagues.
לשפה אחרת
מתוכן המקור
arxiv.org
תובנות מפתח מזוקקות מ:
by Vladimir Som... ב- arxiv.org 04-18-2024
https://arxiv.org/pdf/2404.11335.pdfשאלות מעמיקות