Azadi Soccer Dataset

Azadi soccer dataset is prepared for academic purposes by Image Processing Laboratory (IPL) members of Sharif University of Technology, headed by Mr. Mehran Fotouhi under the supervision of Prof. Shohreh Kasaei, in 2011. The dataset includes the video sequences that are captured from a domestic soccer matches in Iran Pro League. The prepared dataset is one of the few existing multiview soccer datasets which contains 9 fixed cameras and 3 broadcast cameras. The quality of videos is SD, the size of each frame is 720×576 and the frame rate is 25fps. The playfield size and the location of each camera are measured using a set of station theodolites. The dataset contains the actual video capturing phenomena (such as camera lens distortion, motion blur, lighting changes, shadows, small-size multiple moving objects, noise, and so forth). These make it almost ideal for being used as a practical dataset for research purposes. As such, several doctoral and master's theses have been defined at IPL. Some sample videos captured by the broadcast and fixed cameras are shown below. Several sub-datasets are also prepared for different research topics including ball detection, frame synchronization, player detection, and billboard ads. For more information please contact skasaei[at]sharif[dot]edu.


F42

P2center, longview

F22

P2offside

F62

For citation purposes, you can use our publications.




Fixed Cameras

In the Azadi soccer dataset, the locations of fixed cameras have been changed at each halftime. In the first halftime, the whole playfield is covered by fixed cameras and in the second halftime, a half of the playfield is covered by all fixed cameras. Therefore, in the second configuration of fixed cameras, the number of views for a specific region of the playfield is more than that in the first configuration.

Fxy(n) denotes: Fixed camera number x, in the yth halftime, from the nth part of the related source video.

Camera No. Camera Info Total No. of Frames Pose Changed from Initial Calibration Synced Frame No.
Cam 1 F12(1) 47141 Yes 30172
Cam 2 F22 78881 No 35976
Cam 3 F32(2) 53609 No 19924
Cam 4 F42 85606 Yes 41975
Cam 5 F52(8) 43660 Yes 21810
Cam 6 F62 85257 Yes 44358
Cam 7 F72 83857 Yes 42695
Cam 8 F82 83015 No 47663
Cam 9 F92 83530 No 489773

To find the location of fixed cameras in each half of soccer game and other information please contact skasaei[at]sharif[dot]edu.


Ball detection

This sub-dataset contains 2237 frames of 9 fixed cameras (totally 20133 frames), for which the coordination of ball at each view is determined manually and saved in a mat file. To reach this dataset and more information please contact skasaei[at]sharif[dot]edu.


F12

F22

F32

F42

F52

F62

F72

F82

F92

Frame synchronization

This sub-dataset is about synchronization of multiple cameras in multiview camera applications. To reach this dataset and more information please contact skasaei[at]sharif[dot]edu.


F32

F32

F32

F42

F42

F42




Broadcast Cameras

The locations of broadcast cameras in Azadi stadium are standard. In the Azadi dataset, the videos of 2 offside cameras and 1 center camera are available as non-directed broadcast cameras. The camera parameters are varying during the match.

Pyl(n) denotes: PTZ camera, in location l, in the yth halftime, from the nth part of the related source video.

Camera Info Synced Start Time
P2center, longview 1h, 11m, 11s
P2offside(3) 0h, 26m, 57s
P2crane(4) 1h, 10m, 45s

To find the location of broadcast cameras in Azadi soccer dataset and other information please contact skasaei[at]sharif[dot]edu.


Player detection

The player detection dataset contains 6 different sequences of center camera. Each sequence has different number of frames which are gathered from different views. To reach this dataset and more information please contact skasaei[at]sharif[dot]edu.


P2center, longview

P1center, longview

P1center, longview

P1center, longview

P1center, longview

P2center, longview

Billboard ads

This sub-dataset, the coordinats of the 4 corners of a advertisement billboards are determined manually for a sequence of center camera. To reach this dataset and more information please contact skasaei[at]sharif[dot]edu.





Publications

  • M. Fotouhi, S. Fouladi, and S. Kasaei. Projection Matrix by Orthogonal Vanishing Points. Springer, Multimedia Tools & Applications, 2016.
  • Bozorgpour, M. Fotouhi, and S. Kasaei. Robust Homography Optimization in Soccer Scenes. In The 23rd Iranian Conference on Electrical Engineering (ICEE), pages 787-792, Tehran, Iran, May 2015. [BibTeX]
  • M. Fotouhi, A. Bozorgpour, and S. Kasaei. Automatic Soccer Field Line Recognition by Minimum Information. In International Symposium on Artificial Intelligence and Signal Processing (AISP), pages 136-142, Mashhad, Iran, March 2015. [BibTeX]