User’s Guide

The GBU (Good, Bad and Ugly) database consists of parts of the MBGC-V1 image set. It defines three protocols, i.e., Good, Bad and Ugly for which different model and probe images are used.

Installation

To be able to use this database, please have a look at the NIST webpage: http://www.nist.gov/itl/iad/ig/focs.cfm and download The Multiple Biometric Grand Challenge (MBGC)-V1 image database.

Unfortunately, the directory structure in this image database changed. If you have an older version of it, and the bob_dbmanage.py gbu checkfiles --directory <YOUR_PATH_TO_MBGC-V1> fails (i.e. reports missing files), you have two possible options:

  • Call:

    $ bob_dbmanage.py gbu create --recreate --rescan-image-directory <YOUR_PATH_TO_MBGC-V1>

    (you might need root access to recreate the database)

  • Copy (or link) the images of the MBGC-V1 database into a directory that has the required directory structure by calling:

    $ bob_dbmanage.py gbu copy-image-files --soft-link --original-image-directory <YOUR_PATH_TO_MBGC-V1> --new-image-directory <NEW_IMAGE_PATH_TO_BE_CREATED>

To be sure that the procedure succeeded, please call bob_dbmanage.py gbu checkfiles --directory <YOUR_PATH_TO_MBGC-V1> or bob_dbmanage.py gbu checkfiles --directory <NEW_IMAGE_PATH_TO_BE_CREATED> afterwards. If this fails again, your copy of the MBGC-V1 database is corrupted, and you might consider to get a new copy of it.

The Database Interface

The bob.db.gbu.Database complies with the standard biometric verification database as described in bob.db.base <bob.db.base>, implementing the interface bob.db.base.SQLiteDatabase.

Particularities of this Database Interface

This implementation of the GBU database wrapper provides two different types of protocols:

  • ‘gbu’: In the original protocol models are extracted for each file. This means that there may be several models for the same client.

  • ‘multi’: The new protocol provides one or more files for each model. Hence, only one model per client is generated.

There are slight differences between the two protocol types:

  1. The amount of models differs, and so the amount of scores does.

  2. The model_ids differ. For a 'multi' protocol, model_id and client_id are the same, whereas for a 'gbu' protocol, model_id and file_id are identical.

  3. Only the results generated by the 'gbu' protocol are comparable to results that other groups reported on the GBU database.

Note

To be consistent with results published in literature, the default protocol type is 'gbu'.

Warning

Do never mix 'gbu' and 'multi' protocol types. When you are lucky, you will get an exception, if not, other weird things might happen without you noticing!

The three protocols 'Good', 'Bad' and 'Ugly' are provided for both protocol types. Additionally, there are four different training sets (sub-worlds), called 'x1', 'x2', 'x4' and 'x8' with different numbers of training images (the number of training identities is the same for all training sets).

The GBU protocols do only provide a 'dev' set, but no 'eval' set. The default measurement reported is the ROC curve on the ‘dev’ set. When a single number is required, usually the FRR (or the CAR) at FAR=0.1% is reported by other researchers.