Videos

You will find below all the videos you need to evaluate your background subtraction algorithm.

Learning phase

To tune the parameters of your algorithm, you should use the following sequences, generated thanks to the SiVic software, a research version of the pro software Pro-Sivic provided by the Civitec society. You have to download the ground truth of each video, where background pixel are set to 0, and foreground to any other value. In those videos, two scenes are presented: a street and a rotary, and the videos are structured like this: 10 seconds withtout objetcs, then 50 seconds with moving objects. 

Street

  • First configuration (111): Cloudy without acquisition noise, PNG format
  • Second configuration (211): Cloudy with acquisition noise, PNG format
  • Third configuration (311): Sunny with acquisition noise, PNG format
  • Fourth configuration (411): Cloudy, foggy, with acquisition noise, PNG format
  • Fifth configuration (511): Cloudy, windy with acquisition noise, PNG format

Rotary

  • First configuration (121): Cloudy without acquisition noise, PNG format
  • Second configuration (221): Cloudy with acquisition noise, PNG format
  • Third configuration (321): Sunny with acquisition noise, PNG format
  • Fourth configuration (421): Cloudy, foggy, with acquisition noise, PNG format
  • Fifth configuration (521): Cloudy, windy with acquisition noise, PNG format

Evaluation phase

To have a complete evaluation of your algorithms, you have to download a first dataset of synthetic videos, as in the Learning phase. Now, the scenario of the videos is the following: 20 seconds without event, 20 seconds with an event (sun uprising for example), and finally 20 seconds without event. Please note that the ground truth is encrypted with a random process.

Street

  • First configuration (112): Cloudy without acquisition noise, PNG format
  • Second configuration (212): Cloudy with acquisition noise, PNG format
  • Third configuration (312): Sunny with acquisition noise, PNG format
  • Fourth configuration (412): Cloudy, foggy, with acquisition noise, PNG format
  • Fifth configuration (512): Cloudy, windy with acquisition noise, PNG format

Rotary

  • First configuration (122): Cloudy without acquisition noise, PNG format
  • Second configuration (222): Cloudy with acquisition noise, PNG format
  • Third configuration (322): Sunny with acquisition noise, PNG format
  • Fourth configuration (422): Cloudy, foggy, with acquisition noise, PNG format
  • Fifth configuration (522): Cloudy, windy with acquisition noise, PNG format


You have also to test your algorithms with "real" videos. This dataset has been built in order test the algorithms reliability during time and in difficult situations such as outdoor scenes. So, real long videos (about one hour) are available that can present long time change in luminosity with small density of objects in time compared to previous synthetic ones.
This dataset allows test the influence of some difficulties (different ground type, presence of vegetation, casted shadows, presence of a continuous car flow near to the surveillance zone, etc.) encountered during the object extraction phase. Please note that the ground truth is encrypted with a random process.

Real applications


NEW: context information for real video 1!

You will find below all the information about the acquisition of Video 001. 
  • Focal length: 9mm
  • Size of CCD: 6.4x4.8 mm²
  • Resolution: 320x240 pixels
  • GPS coordinates: 45.75381N 4.92536E
  • Height of the camera: 5.8m
  • Pan camera: +126° (0° for North, positive for West)
  • Tilt camera: -6° (0° for horizon, positive for looking up)
  • Roll camera: -5° (positive for clock-wise direction)