Bachelor's Thesis: Low-Resolution Face Recognition using Rank Lists

This repository serves as code base for the corresponding bachelor's thesis. The code is written in Python v3.7 and uses bob.bio.face.

Main packages: bob : {bio.base, bio.face, pipelines}, sklearn : {pipeline}, numpy, scipy, argpars, pandas

Installation / Virtual Environment (Linux)

1. First, install mambaforge.
2. Configure the conda installation by executing the following commands:

conda install -n base -c conda-forge mamba
mamba update -n base -c conda-forge conda mamba
conda config --set show_channel_urls True

3. Create a virtual environment using mamba as follows:

mamba create --name bob_env --override-channels -c https://www.idiap.ch/software/bob/conda -c conda-forge python=3 bob.io.base bob.bio.face pytorch  

3. Activate the environment by conda activate bob_env and configure its channels as follows:

conda config --env --add channels conda-forge
conda config --env --add channels https://www.idiap.ch/software/bob/conda

4. Install packages by using mamba install <package-name>.
5. Download and save the SCface database.
6. In the activated environment, configure Bob with the path to the database directory as follows:

bob config set bob.bio.face.scface.directory <path>

7. The configuration can be checked using:

bob config show

Pipeline Setup

To successfully install the required packages for running the simple_pipe.py, follow these steps:

1. Install buildout.cfg pip install zc.buildout.
2. Run buildout. This should create a couple of folders and files.
3. Whenever bob or python is called through the terminal, use bin/bob and bin/python instead.

Feature Extraction

Install all necessary packages for simple_pipe.py and run the following command three times using close, medium, and far each once as protocol.

bin/bob bio pipelines vanilla-biometrics scface-<protocol> ./helpers/simple_pipe.py -vvv -o samples_pipe_all -c --group eval

This runs the pipeline and saves all checkpoint data in a folder called samples_pipe_all. The extracted features used for this project can be found in the subdirectory called samplewrapper-2. To read the files, use the terminal with the command h5dump -y <filename>.h5

Running the Code

Use the terminal for starting the script with the command bin/python main.py. The following options are available:

• --protocol, -p
  Choose between the close, medium, or far protocol used for the comparison (default: all)
• --comparison_method, -c
  Specify a comparison method (default: all)

Approach	Comparison Methods
direct comparison	baseline
rank list comparison (use rank_list_comparison to run all methods)	mueller2010, schroff, mueller2013, wartmann, spearman, kendall, weighted_kendall
standardization comparison (use standardization_comparison to run all methods)	cosine, braycurtis, canberra, cityblock, sqeuclidean, minkowski

• --standardization_method, -s
  Select standardize, subtract_mean, or omitted to define the preprocessing of the lists with cosine distances. This argument is only important whenever methods for a standardization comparison are chosen (default: standardize).
• --record_output, -r
  Include to record scores such as recognition rates, score files, preprocessing time, and runtime (default: False)
• --enable_larger_cohort, -lc
  Include to extend the cohort with 43 samples (default: False)

Evaluation of Results

If the recording of scores is enabled, an output directory is created in which there are .csv files for every comparison method and protocol used for verification, as well as a .csv file containing the recognition rates used for identification. To evaluate the verification results, use the following command, including the files for visualizing:

bin/bob bio roc -v -o <output-filename>.pdf -ts "<graph-title>" -lg <line-names> <csv-files>

The repository additionally includes an example_output directory which includes plots and recorded output. Furthermore, the plots repository includes python executables for the comparison methods wartmann and minkowski, which generate plots with values computed with various parameters.