ICTEAM > ELEN > ISPGroup > APIDIS


  ISPGroup

> Welcome

> Members

> Research

> Publications

> Seminars

> Projects

> Mailing Lists

> Data/Software

> Open Positions

> (login)

  Sponsors
















APIDIS basket ball dataset

! APIDIS basket ball dataset

This page gives access to the first acquisition campaign of basket ball data during the APIDIS European project.

A second acquisition campaign (the SPIROUDOME dataset) is available here.

Please scroll down for the detections and the ground-truth of players and ball.

Terms of use: [pdf] This dataset is available for non-commercial research in video signal processing only. We kindly ask you to mention the APIDIS project when using this dataset (in publications, video demonstrations...).

Acknowledgements

We would like to thank Jean-François Prior (Dexia Namur basket ball team), Philippe Delmulle (Declercq Stortbeton Waregem basket ball team) and the city of Namur for their authorisations and technical help collecting this dataset.

The dataset is composed of a basket ball game:

  • Seven 2-Mpixels color cameras around and on top of a basket ball court
  • Time stamp for each frame
All cameras being captured by a unique server at ~22 fps
  • Manually annotated basket ball events for the entire game
These metadata provide events relative to the basket ball game, e.g. ball possession periods, throws, violations (see NEM'08 summit paper).
  • Manually annotated objects positions for one minute of the game
These metadata provide the positions of the players, referees, baskets and ball in each frame of each camera (see NEM'08 summit paper).
  • Calibration data
These are measures of the basket ball court and calibration pictures that can be used for calibrating each camera into a common world coordinate system.

Note: Due to bandwidth limitations, only a part of the basket ball game is available from this web site. Please contact us (bottom of this page) for more data.

The following table provides the dataset organised in the three following sections:

  • Common data
Data that are common to all video cameras, e.g. basket ball court measures, annotated basket ball events, files formats of the dataset.
  • Original dataset
For each camera, the video files, annotated objects positions and some calibration pictures are provided.
  • Pseudo synchronized dataset
This is a dataset that provides synchronized 25 fps video streams for all cameras. It has been computed from the original dataset. It is provided for research in joint video processing algorithms but the time stamps are less accurate than in the original dataset.

 

Common data

Calibration data

Calibration

Manually annotated basket events

Events

Hardware

Cameras: All cameras are Arecont Vision AV2100M IP cameras. The datasheets can be downloaded from the constructor site here.
Lenses: The fish-eye lenses used for the top view cameras are Fujinon FE185C086HA-1 lenses. The datasheet can be downloaded from the constructor site here.

Time stamps

All time stamps are expressed in terms of seconds since Epoch when provided as integers. When provided in a human readable format, e.g. in filenames, they follow the ISO 8601 date/time syntax.

Files format

  • Video files
    • The video files for each camera are available in their native format, i.e. one motion jpeg file per minute per camera (almost 300 MB per minute per camera). The size is 1600x1200 and the framerate is almost 22 fps in average.
    • In the pseudo synchronised dataset below, the video files are recorded at 25 fps in 800x600 resolution in MPEG-4. Their size is between 28 MB and 56 MB for one minute.
  • Time stamps
    • Index files
    A .idx file is associated with each video file. It provides the accurate time stamp for each frame.
    The structure of those .idx files is as follows:
    • Header
      • unsigned 64 bits header (cookie)
      • unsigned 32 bits header (index file version)
    • For each record (24 bytes):
      • unsigned 32 bits integer (number of seconds since Epoch)
      • unsigned 32 bits integer (number of microseconds, i.e. less than 1,000,000)
      • unsigned 64 bits integer (frame offset in bytes in the video file)
      • unsigned 32 bits integer (frame number in the file)
      • unsigned 32 bits integer (reserved)
    All integers are saved in LITTLE_ENDIAN byte order.
    In the case of mjpeg video files, this index file allows to extract one jpeg frame (e.g. with libjpeg) from the .mjpeg file and provides its time stamp.
  • Metadata XML files
    A simplified structural diagram of event xml files is: event-xml-simple.png.
    You can also find a full view of all tags defined in apidis-annotation-ver23.xsd and their structures here.
    The following diagram shows the tags for describing the detected objects and their properties: salient-obj-xml.png.

 

Original dataset

Camera thumbnail

Camera index

7

4

2

5

3

1

6

Video files and time stamps

Video files and time stamps

Video files and time stamps

Video files and time stamps

Video files and time stamps

Video files and time stamps

Video files and time stamps

Video files and time stamps

Manually annotated objects positions

Objects positions
(one minute only)

Objects positions
(one minute only)

Objects positions
(one minute only)

Objects positions
(one minute only)

Objects positions
(one minute only)

Objects positions
(one minute only)

Objects positions
(one minute only)

Calibration pictures

Calibration pictures

Calibration pictures

Calibration pictures

Calibration pictures

Calibration pictures

Calibration pictures

Calibration pictures

 

Pseudo synchronised dataset

 

The pseudo synchronised dataset contains only one minute of the game. Using an exact frequency of 25 Hz, this dataset uses the closest acquired frame (temporally) and its associated metadata. Therefore, this dataset provides only the best 25 fps approximation we can have from the original dataset. For optimal time stamps accuracy, the original dataset should be prefered. The dataset is available here.

A Matlab code is available here to access the video frames.

Ball detection and tracking ground truth (available here), as used in:

  • Parisot, Pascaline ; De Vleeschouwer, Christophe. Graph-based filtering of ballistic trajectory. IEEE International Conference on Multimedia and Expo (ICME), Barcelona, July 2011.
  • K.C., Amit Kumar ; Parisot, Pascaline ; De Vleeschouwer, Christophe. Spatio-temporal template matching for ball detection. ACM/IEEE International Conference on Distributed Smart Cameras (ICDSC), Ghent, Belgium, August 2011.
  • Parisot, Pascaline; De Vleeschouwer, Christophe. "Consensus-based trajectory estimation for ball detection in a calibrated cameras system." (submitted to the Journal of Real-Time Image Processing) (code available here)

Player detection and tracking ground truth & evaluation metrics (available here), as used in:

As the detection accuracy is intrinsically limited due to calibration and synchronization errors affecting the dataset, we recommend to use 30 cm as the threshold value for computing the MOTA metric.

  • Delannay, Damien ; Danhier, Nicolas ; De Vleeschouwer, Christophe. Detection and recognition of sports(wo)men from multiple views. ACM/IEEE International Conference on Distributed Smart Cameras (ICDSC), Como, Italy, August 2009.
  • K.C., Amit Kumar ; De Vleeschouwer, Christophe. Discriminative Label Propagation for Multi-Object Tracking with Sporadic Appearance Features. International Conference on Computer Vision (ICCV), Sydney, Australia, December 2013
  • Parisot, Pascaline ; Sevilmis, Berk ; De Vleeschouwer, Christophe. Training with corrupted labels to reinforce a probably correct teamsport player detector. Advanced Concepts for Intelligent Vision Systems. Lecture Notes in Computer Science Volume 8192, 2013. (pdf)
  • Parisot, Pascaline ; De Vleeschouwer, Christophe. Scene-specific classifier for effective and efficient team sport players detection from a single calibrated camera (in submission at CVIU - Special Issue on Computer Vision in Sports) (databases available here)

Related publications:

  • Chen, Fan ; De Vleeschouwer, Christophe. Personalized production of basketball videos from multi-sensored data under limited display resolution. Computer Vision and Image Understanding, Vol. 114, no. 6, p. 667-680, 2010.
  • De Vleeschouwer, Christophe ; Chen, Fan ; Delannay, Damien ; Parisot, Christophe ; Chaudy, Christophe ; Martrou, Eric ; Cavallaro, Andrea. Distributed video acquisition and annotation for sport-event summarization, NEM Summit, Saint Malo, France, 2008.
  • Chen, Fan ; Delannay, Damien ; De Vleeschouwer, Christophe. An Autonomous Framework to Produce and Distribute Personalized Team-Sport Video Summaries: A Basketball Case Study, IEEE Transactions on Multimedia, Volume:13 , Issue: 6, pp. 1381 - 1394, December 2011.

If necessary, please contact the coordinator: Christophe De Vleeschouwer, christophe.devleeschouwer@uclouvain.be. You can also contact Damien Delannay, Damien.Delannay@uclouvain.be.


 



Last updated July 29, 2016, at 01:12 PM