SoccerNet Calibration

Camera calibration is the link between the image world and the 3D real world. Automatic calibration of the camera is an important topic of research for sports analytics that can lead to interesting applications such as offline line analysis. It is also the key to integrate reality graphics into any live production. We defined two tasks on this topic: soccer pitch marking and goal post localization, and the automatic camera calibration, with the results of the first task that can be used as inputs to the second.

First task: Soccer pitch markings and goal posts localization

Our task.

Given an image, detect the extremities of every soccer pitch element captured in the image. A soccer pitch element is either a pitch line marking or a goal post part. The extremities are either the end of the line/circle arc or its intersection with the image side. Since the notion of extremities for the "Circle central" semantic class is not well-defined, we will not use this class in the evaluation. However, note that still detecting it might be useful for the second task.

All of our classes.

We provide annotation for all soccer lines and goal posts: {Big rect. left bottom, Big rect. left main, Big rect. left top, Big rect. right bottom, Big rect. right main, Big rect. right top, Circle central, Circle left, Circle right, Goal left crossbar, Goal left post left , Goal left post right, Goal right crossbar, Goal right post left, Goal right post right, Goal unknown, Line unknown, Middle line, Side line bottom, Side line left, Side line right, Side line top, Small rect. left bottom, Small rect. left main, Small rect. left top, Small rect. right bottom, Small rect. right main, Small rect. right top}.

Our data.

The data consists of 20,028 images taken from the SoccerNet videos at events annotated for the action spotting task, and images from their replays. The challenge set is composed of 2,104 separate images from different games.

Our Metric.

The evaluation is indirectly based on the euclidean distance between the predicted extremities and the annotated extremities in the image resized in 1/4 HD (960,540). Rather than using distance metrics in pixels explicitly which are often more difficult to interpret, we formulate the evaluation using accuracy, precision and recall as for a binary classification problem.

Second task: Automatic camera calibration

Our task.

Given a common 3D pitch template, the camera parameters are used to estimate the reprojection error induced by the camera parameters. The camera parameters include its lens parameters, its orientation, its translation with respect to the world reference axis system that we define accordingly.

Our data.

The data consists of the same 20,028 images as for the first task, so taken from the 500 games of soccernet at events annotated for the action spotting task, and images from their replays. Likewise, The challenge set is composed of the same 2,104 images.

Our Metric.

The evaluation is based on the reprojection error which we define here as the L2 distance between one annotated point and the line to which the point belong. We consider a pitch marking to be one entity, and for it to be correctly detected, all its extremities (or all points annotated for circles) must have a reprojection error smaller than the threshold. We also measure the completeness rate as the number of camera parameters provided divided by the number of images with more than four semantic line annotations in the dataset.

For more details, check out our development kit on github

Our videos on Calibration

Soccer Player Tracking, Re-ID, Camera Calibration and Action Spotting - SoccerNet Challenges 2022

In this video, we present our new SoccerNet Challenges for CVPR 2022! We introduce the three tasks of Calibration, Re-identification and Tracking on soccer games, in partnership with EVS Broadcast Equipment, SportRadar and Baidu Research. We also reiterate our previous Action Spotting and Replay Grounding Challenges at the ActivityNet workshop.

How to cite this work ?

available soon