# Evaluating deep tracking models for player tracking in broadcast ice hockey video

**Authors:** Kanav Vats, Mehrnaz Fani, David A. Clausi, John S. Zelek  
**arXiv:** [2205.10949](https://arxiv.org/abs/2205.10949)  
**Date:** 2022-05-22  
**Summary page:** https://www.summarizepaper.com/en/paper/2205.10949v1/

## Key points

- The paper addresses the challenge of tracking and identifying players in computer vision-based ice hockey analytics.
- Prior research has used handcrafted features for player detection, but lacks publicly reported network architectures, training data, and performance metrics for commercial solutions in hockey player tracking.
- The researchers compare state-of-the-art tracking algorithms to analyze their performance and failure modes specifically within the context of ice hockey.
- The study aims to provide insights into the effectiveness of different tracking models for accurately identifying and following players during gameplay.
- The authors detail their annotation process for creating a comprehensive hockey player tracking dataset, with 15 annotators using the CVAT tool to annotate 84 videos.
- The researchers present comparisons between their tracking dataset and other multi-object datasets commonly used in literature such as MOT16, MOT20, and KITTI-T.

## Summary

In their paper "Evaluating deep tracking models for player tracking in broadcast ice hockey video," Kanav Vats, Mehrnaz Fani, David A. Clausi, and John S. Zelek address the crucial issue of tracking and identifying players in computer vision-based ice hockey analytics. The fast-paced and non-linear motion of players in hockey poses a significant challenge for player tracking, exacerbated by player-player and player-board occlusion, camera panning, and zooming in hockey broadcast videos. While prior research has utilized handcrafted features for player detection and re-identification, the authors note a lack of publicly reported network architectures, training data, or performance metrics for commercial solutions in hockey player tracking. The researchers highlight the absence of published work on hockey player tracking leveraging recent advancements in deep learning techniques while also reporting accuracy metrics found in existing literature. To fill this gap, they compare and contrast several state-of-the-art tracking algorithms to analyze their performance and failure modes specifically within the context of ice hockey. The study aims to provide insights into the effectiveness of different tracking models for accurately identifying and following players during gameplay. Additionally, the authors provide detailed information on their annotation process for creating a comprehensive hockey player tracking dataset. Fifteen annotators used the CVAT tool to annotate 84 videos with tight bounding boxes around players/referees, employing linear interpolation to interpolate bounding box positions. The average time taken to annotate one minute of video was 10.45 minutes, totaling 527 minutes for all videos. Furthermore,the researchers present comparisons between their tracking dataset and other multi-object datasets commonly used in literature such as MOT16,MOT20,and KITTI-T.They showcase the scale of their dataset with 84 videos containing 91,and 773,related to ice hockey players. Overall, this study contributes valuable insights into the challenges and advancements in using deep learning models while providing a detailed methodology for dataset creation and evaluation.