Welcome to the SoccerNet Development Kit for the Tracking Task and Challenge. This kit is meant as a help to get started working with the data and the proposed task. More information about the dataset can be found on our official website.
The Tracking dataset consists of 12 complete soccer games from the main camera including:
- 200 clips of 30 seconds with tracking data.
- one complete halftime annotated with tracking data.
- the complete videos for the 12 games.
Note that a subset of this data is used in this first challenge. In particular, this accounts for 57 30-seconds clips for the train set, 49 clips for the test set, 58 clips for our first public challenge, and 37 clips for our future challenges, including the entire half-time video in the latter.
Participate in our upcoming Challenge at CVPR and try to win up to 1000$ sponsored by Baidu! All details are available on the challenge website, or on the main page.
The participation deadline is fixed at the 30th of May 2022. The official rules and guidelines are provided in ChallengeRules.md.
We provide a SoccerNet pip package to easily download the data and the annotations.
To install the pip package simply run:
pip install SoccerNet
Then, to download the tracking data, enter the following commands:
from SoccerNet.Downloader import SoccerNetDownloader
mySoccerNetDownloader = SoccerNetDownloader(LocalDirectory="path/to/SoccerNet")
mySoccerNetDownloader.downloadDataTask(task="tracking", split=["train","test","challenge"])The ground truth and detections are stored in comma-separate csv files with 10 columns. These values correspond in order to: frame ID, track ID, top left coordinate of the bounding box, top y coordinate, width, height, confidence score for the detection (always 1. for the ground truth) and the remaining values are set to -1 as they are not used in our dataset, but are needed to comply with the MOT20 requirements.
Multi-object tracking (MOT) aims at recovering trajectories of multiple objects in time by estimating object bounding boxes and identities in videos sequences. We consider two tasks: (1) a pure association task that considers ground-truth detections (task of the first challenge), or (2) a complete tracking task that expects detecting the objects of interest from the raw video (task for benchmarking and future challenges).
So for the first challenge, one may use the provided ground-truth bounding boxes and focus only on the association task.
The object classes are not taken into account in this challenge or the evaluation. The object to retrieve are among the following classes: players, goalkeepers, referees, balls and any other human entering the field.
For our benchmark and challenge, we consider HOTA as the main metric. More specifically, this metric can be decomposed into two components: DetA and AssA, focusing on detection and association accuracy, respectively.
This repository contains several benchmarks for the tracking task. You can use these codes to build upon state-of-the-art methods and improve the performances.
We use the TrackEval repository to evaluate the performances. To participate the challenge, you should submit the result as a zip file to EvalAI. Please see the READMEs of our baseline methods under the "Benchmarks" directory for instructions on how to prepare the zipped results.
You can use the MOT Challenge Evaluation Kit to visualize the tracks.
Check out our other challenges related to SoccerNet!
For further information check out the paper and supplementary material:
Please cite our work if you use our dataset:
Available soon