home   |  research   |  members   |  projects   |  publications   |  conferences of interest   |  downloads   |  contact & impressum   |  privacy information

AGESEG -- Casia Ageing Database Segmentation Ground Truth

This is "The Multimedia Signal Processing and Security Lab", short WaveLab, website. We are a research group at the Artificial Intelligence and Human Interfaces (AIHI) Department of the University of Salzburg led by Andreas Uhl. Our research is focused on Visual Data Processing and associated security questions. Most of our work is currently concentrated on Biometrics, Media Forensics and Media Security, Medical Image and Video Analysis, and application oriented fundamental research in digital humanities, individualised aquaculture and sustainable wood industry.


This is a manual segmentation of a subset of the Project Peking (Casia) Subject Aging Database. The superset database of iris images is published as CASIA Iris Subject Ageing Version 1.0 at idealtest.

In this database only the ground truth is made available in the form of text files containg the boundary points. These files, along with the original image, can be used with manuseg from the USIT to generate normalized templates and masks.

The segmentation was done by a single operator OperatorA, the relevant files are in the OperatorA subdirectory.

The database and first findings are described in more detail in

Heinz Hofbauer, Inmaculada Tomeo-Reyes, and Andreas Uhl, “Isolating Iris Template Ageing in a Semi-controlled Environment”, in Proceedings of the International Conference of the Biometrics Special Interest Group, 2016.

Selection of users

Adhering to the higher quality sessions according to:

Peter Wild, James Ferryman, and Andreas Uhl, “Impact of (segmentation) quality on long vs. short-timespan assessments in iris recognition performance”, IET Biometrics 4:4, pp. 227-235, IET, 2015. DOI: 10.1049/iet-bmt.2014.0073

we selected the following sessions:

  • 2009: Session 1, Session is without glasses
  • 2013: Without Glasses

Some users had to be removed from the segmentation:

  • 0023, 1067 no data for user in 2009 (empty directory)
  • 1004 no images without glasses in 2013
  • 0190 was in user 0189 directory in 2009, no data in 2013
  • 0191 was in user 0189 directory in 2009, no data in 2013

The same happened with user 0000 in 2009, the file id was 0001. This was corrected in the selected set.

Original images:

  • 2009: 120 images from an video sequence per eye (total of 240 per user)
  • 2013: 20 images from a video sequence per eye (total of 40 per user)

Subset selection

We choose ten images per eye spaced apart as far as possible in the temporal domain.

Note that some images are temporally closer than others, this was done when blinking in the video sequence would obfuscate large parts of the iris and larger temporal spacing could not be maintained.

Total of 47 users and with left and eye counted separately 94 different eyes, total of 470 images per year.

File makeup

  • Year: the year of acquisition
  • Sens: Sensor ID 0281 is the Irisguard H100
  • User: User id
  • Eye: left (L) or right (R) eye
  • Idx: Running index for the user in this session (lower index should be earlier time)
  • Session: Recording session: session 1 was used from the 2009 data; 2013 only had one session with id 0
  • Unknown
  • Boundary: The boundary: inner and outer for pupillary and limbic boundary (fit ellipse) as well as upper and lower for eyelids (fit polynomial).

Tip for usage with caht

Since caht by default uses some fixed parameters it is beneficial to crop the images for caht application.

convert -crop 440x380+100+50

Sources and Executables

The code is available upon request.

Please fill out this form to request a download link for the software:

Email address: