xfacereclib.paper.IET2014 1.0.0

Important!

After downloading the databases, you will need to tell our software, where it can find them by changing the configuration files. In particular, please update the scface_directory in xfacereclib/paper/IET2014/database_scface.py, as well as mobio_image_directory and mobio_annotation_directory in xfacereclib/paper/IET2014/database_mobio.py. Please let all other configuration parameters unchanged as this might influence the face recognition experiments and, hence, the reproducibility of the results.

Generating output files

Also, all information are written to console (when using the -vvv option to enable debug information), including:

1. The \(C^{\mathrm{min}}_{\mathrm{ver}}\) of the development set, the \(C^{\mathrm{min}}_{\mathrm{ver}}\) of the evaluation set and the \(C_{\mathrm{ver}}\) of the evaluation set based on the optimal threshold on the development set.

2. The \(C_{\mathrm{frr}}\) on both development and evaluation set, using the threshold defined at FAR=1% of the development set.

3. The \(C_{\mathrm{ver}}\) on the development and evaluation set, when applying threshold \(\theta_0=0\) (mainly useful for calibrated scores).

4. The \(C_{\mathrm{cllr}}\) performance on the development and the evaluation set.

5. The \(C^{\mathrm{min}}_{\mathrm{cllr}}\) performance on the development and the evaluation set.

All these numbers are computed with and without ZT score normalization, and before and after score calibration.

To run the script, some command line parameters can be specified, see ./bin/iec2014_evaluate.py --help:

• --result-directory: Specify the directory where final result files are stored (default: results). This should be the same directory as passed to the bin/iec2014_execute.py` script.

• --databases: Specify a list of databases that you want evaluate. Possible values are scface and mobio. By default, both databases are evaluated.

• --protocols: Specify a list of protocols that you want to evaluate. Possible values are combined, close, medium and far for database scface, and male and female for mobio. By default, all protocols are used.

• --combined-zt-norm: Evaluate the face recognition experiments on the SCface database with combined ZT-norm cohort.

• --combined-threshold: Evaluate the face recognition experiments on the SCface database by computing the threshold on the combined development set.

• --latex-directory: The directory, where the final score files will be placed into, by default this directory is latex.

Again, the most usual way to compute the resulting tables could be:

1. Evaluate experiments on MOBIO:

   $ bin/iet2014_evaluate.py -vvv --database mobio

2. Evaluate experiments on SCface with distance-dependent ZT-norm:

   $ bin/iet2014_evaluate.py -vvv --database scface

3. Evaluate experiments on SCface with distance-independent ZT-norm:

   $ bin/iet2014_evaluate.py -vvv --database scface --combined-zt-norm --protocols close medium far

4. Evaluate experiments on SCface with distance-independent threshold (will mainly change the \(C_{\mathrm{ver}}\) of the evaluation set):

   $ bin/iet2014_evaluate.py -vvv --database scface --combined-threshold --protocols close medium far

5. The experiments to compare linear calibration with categorical calibration as given in Table 7 of the Paper are run using the bin/iet2014_categorical.py script:

   $ bin/iet2014_categorical.py -vvv

Generate the LaTeX tables

Finally, the LaTeX tables can be regenerated by defining the accordant \Result and \ResultAtZero LaTeX macros and include the resulting files. E.g., to create Table 3 of the Paper, define:

\newcommand\ResultIII[2]{\\}
\newcommand\ResultII[9]{#1\,\% \ResultIII}
\newcommand\Result[9]{#1\,\% & #4\,\% & #2\,\% & #3\,\% & #5\,\% & #6\,\% & #9\,\% & #7\,\% & #8\,\% &\ResultII}
\newcommand\ResultAtZero[8]{}

set up your tabular environment with 10 columns and input at according places:

\input{latex/mobio_male}
\input{latex/mobio_female}
\input{latex/scface_close}
\input{latex/scface_medium}
\input{latex/scface_far}
\input{latex/scface_combined}

Accordingly, the other tables can be generated from files:

• Table 4a): latex/scface_close-zt.tex, latex/scface_medium-zt.tex and latex/scface_far-zt.tex

• Table 4b): latex/scface_close-thres.tex, latex/scface_medium-thres.tex and latex/scface_far-thres.tex

• Tables 5 and 6: latex/mobio_male.tex, latex/mobio_female.tex, latex/scface_close-zt.tex, latex/scface_medium-zt.tex, latex/scface_far-zt.tex and latex/scface_combined.tex .

• Table 7: latex/calibration-none.tex, latex/calibration-linear.tex and latex/calibration-categorical.tex

Generate the score distribution plots

At the end, also the score distribution plots that are shown in Figures 3 and 4 of the Paper can be regenerated. These plots require the face recognition experiments to have finished, and also the categorical calibration to have run. Afterwards, the script bin/iet2014_plot.py can be executed. Again, the script has a list of command line options:

• --result-directory: Specify the directory where final result files are stored (default: results). This should be the same directory as passed to the bin/iec2014_execute.py` script.

• --figure: Specify, which figure you want to create. Possible values are 3 and 4.

• --output-file: Specify the file, where the plots should be written to. By default, this is Figure_3.pdf or Figure_4.pdf for --figure 3 or --figure 4, respectively.

Hence, running:

$ ./bin/iet2014_plot.py -vv --figure 3
$ ./bin/iet2014_plot.py -vv --figure 4

should be sufficient to generate the plots.