Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

SoccerNet-Tracking: Multiple Object Tracking Dataset and Benchmark in Soccer Videos

Anthony Cioppa, Silvio Giancola, Adrien Deliege, Le Kang, Xin Zhou, Zhiyu Cheng, Bernard Ghanem, Marc Van Droogenbroeck

TL;DR

SoccerNet-Tracking tackles the lack of soccer-specific MOT data and benchmarks by introducing a large-scale, single-camera dataset with dense annotations for players, referees, and the ball. The authors provide 200 clips of 30s and a 45-minute half-time video, split into train/test and two private challenges, enabling comprehensive short- and long-term tracking evaluation. They benchmark state-of-the-art MOT methods (DeepSORT, FairMOT, ByteTrack) and demonstrate that long-term re-identification and dense occlusions remain challenging, with fine-tuning improving results. This dataset enables soccer analytics such as pass sequences, formations, and possession from tracking data, and invites future research into robust re-id and long-horizon tracking.

Abstract

Tracking objects in soccer videos is extremely important to gather both player and team statistics, whether it is to estimate the total distance run, the ball possession or the team formation. Video processing can help automating the extraction of those information, without the need of any invasive sensor, hence applicable to any team on any stadium. Yet, the availability of datasets to train learnable models and benchmarks to evaluate methods on a common testbed is very limited. In this work, we propose a novel dataset for multiple object tracking composed of 200 sequences of 30s each, representative of challenging soccer scenarios, and a complete 45-minutes half-time for long-term tracking. The dataset is fully annotated with bounding boxes and tracklet IDs, enabling the training of MOT baselines in the soccer domain and a full benchmarking of those methods on our segregated challenge sets. Our analysis shows that multiple player, referee and ball tracking in soccer videos is far from being solved, with several improvement required in case of fast motion or in scenarios of severe occlusion.

SoccerNet-Tracking: Multiple Object Tracking Dataset and Benchmark in Soccer Videos

TL;DR

SoccerNet-Tracking tackles the lack of soccer-specific MOT data and benchmarks by introducing a large-scale, single-camera dataset with dense annotations for players, referees, and the ball. The authors provide 200 clips of 30s and a 45-minute half-time video, split into train/test and two private challenges, enabling comprehensive short- and long-term tracking evaluation. They benchmark state-of-the-art MOT methods (DeepSORT, FairMOT, ByteTrack) and demonstrate that long-term re-identification and dense occlusions remain challenging, with fine-tuning improving results. This dataset enables soccer analytics such as pass sequences, formations, and possession from tracking data, and invites future research into robust re-id and long-horizon tracking.

Abstract

Tracking objects in soccer videos is extremely important to gather both player and team statistics, whether it is to estimate the total distance run, the ball possession or the team formation. Video processing can help automating the extraction of those information, without the need of any invasive sensor, hence applicable to any team on any stadium. Yet, the availability of datasets to train learnable models and benchmarks to evaluate methods on a common testbed is very limited. In this work, we propose a novel dataset for multiple object tracking composed of 200 sequences of 30s each, representative of challenging soccer scenarios, and a complete 45-minutes half-time for long-term tracking. The dataset is fully annotated with bounding boxes and tracklet IDs, enabling the training of MOT baselines in the soccer domain and a full benchmarking of those methods on our segregated challenge sets. Our analysis shows that multiple player, referee and ball tracking in soccer videos is far from being solved, with several improvement required in case of fast motion or in scenarios of severe occlusion.
Paper Structure (8 sections, 6 figures, 4 tables)

This paper contains 8 sections, 6 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Multiple Object Tracking (MOT)
  4. Sports Video Understanding
  5. Player Localization and Tracking in Sports.
  6. SoccerNet-Tracking Dataset
  7. Benchmarks
  8. Conclusion

Figures (6)

  • Figure 1: SoccerNet-Tracking. We propose a novel dataset for Multiple Object Tracking (MOT) in soccer videos including the players, the ball, and the referees. Our dataset is composed of $200$ sequences of $30$s each, representative of interesting moments from $12$ soccer games, densely annotated with player tracklets, teams and jersey numbers. Moreover, we also include a fully annotated $45$min half time video, focusing on long-term tracking.
  • Figure 2: SoccerNet-Tracking against other tracking datasets. The area is proportional to the number of unique tracklets in each dataset. SoccerNet-Tracking offers a great trade-off between the total number of bounding boxes, frames and unique tracklets. Furthermore, only SSET proposes soccer sequences, which is much smaller and only focuses on single object tracking.
  • Figure 3: Distribution of the action classes. Number of action classes within the $200$ clips and the whole half-time separately. For the $200$ clips, the key action distribution correspond to the anchoring actions in the clip selection process. Note that within the $12$ games, we have no red card or yellow to red card events.
  • Figure 4: Example of tracking annotations in our dataset for challenging events. From top to bottom: (a) Corner actions often display a lot of occluded and clustered players. (b) Direct free-kicks also show clustered players going in the same direction with many crossings between players. (c) Penalties display almost all objects moving in the same direction often followed by cheering. (d) Shot on target actions involve high speed movement of the ball and players towards the goal.
  • Figure 5: Qualitative tracking results. Tracking sequences produced by ByteTrack with ground-truth detections. Sequence (a) represents a good tracking of the players, even after some players or the referee are partially occluded. Sequence (b) shows an example of challenging association due to fast motion of the ball and players between consecutive frames. Sequence (c) displays a challenging free-kick scenario where many players are clustered together resulting in extreme occlusions and poor tracking results.
  • ...and 1 more figures