# D$^{\bf{3}}$: Duplicate Detection Decontaminator for Multi-Athlete Tracking in Sports Videos

**Authors:** Rui He, Zehua Fu, Qingjie Liu, Yunhong Wang, Xunxun Chen  
**arXiv:** [2209.12248](https://arxiv.org/abs/2209.12248)  
**Date:** 2022-09-25  
**Summary page:** https://www.summarizepaper.com/en/paper/2209.12248v1/

## Key points

- Tracking multiple athletes in sports videos is challenging due to similar appearances and occlusion misreporting.
- The authors propose a new definition for duplicate detection and introduce a transformer-based model called D$^3$ for training.
- They also propose a specific matching algorithm called Rally-Hungarian (RH) that is triggered by team sports substitution rules.
- The authors release a new dataset named RallyTrack to complement existing tracking datasets that lack shot changes.
- Combining D$^3$ and RH significantly improves tracking performance with 9.2 in MOTA and 4.5 in HOTA on RallyTrack.
- D$^3$ can accelerate convergence during training, saving up to 80% of the original training time on MOT17.
- The model trained only with volleyball videos can be applied directly to basketball and soccer videos for Multi-Athlete Tracking (MAT).
- The proposed method shows significant improvements over existing approaches and has broad applicability across different types of sports videos beyond those used for training purposes.

## Summary

The task of tracking multiple athletes in sports videos is a challenging one, as the athletes often have similar appearances and are frequently covered by each other, leading to duplicate detection. In this paper, the authors introduce a new definition for duplicate detection as occlusion misreporting on the same athlete by multiple detection boxes in one frame. To address this problem, they propose a novel transformer-based Duplicate Detection Decontaminator (D$^3$) for training and a specific algorithm called Rally-Hungarian (RH) for matching. The D$^3$ model generates enhanced box losses when duplicate detection occurs, while RH is triggered by team sports substitution rules and is particularly suitable for sports videos. To complement existing tracking datasets that lack shot changes, the authors release a new dataset based on sports video named RallyTrack. Extensive experiments on RallyTrack show that combining D$^3$ and RH can significantly improve tracking performance with 9.2 in MOTA and 4.5 in HOTA. Moreover, experiments on MOT-series and DanceTrack demonstrate that D$^3$ can accelerate convergence during training, saving up to 80% of the original training time on MOT17. Interestingly, the authors train their model only with volleyball videos but find that it can be applied directly to basketball and soccer videos for Multi-Athlete Tracking (MAT), demonstrating the priority of their method. The dataset used in this study is available at https://github.com/heruihr/rallytrack. Overall, this paper presents an innovative approach to addressing duplicate detection in multi-athlete tracking tasks using a transformer-based model and specific matching algorithm tailored to team sports substitution rules. Their proposed method shows significant improvements over existing approaches and has broad applicability across different types of sports videos beyond those used for training purposes.