Running baseline experiments and evaluations for the IJCB 2017 UCCS challenge

## Project description

UCCS Face Detection and Recognition Challenge
=============================================

This package implements the baseline algorithms and evaluation for part 2 and 3 of the face recognition challenge.
This package relies on the signal processing and machine learning library Bob_.
For installation instructions and requirements of Bob_, please refer to the Bob_ web page.

.. note::
Due to limitations of Bob_, this package will run only under Linux and MacOS operating systems.
Particularly, it will not work under any version of Microsoft Windows, and maybe not under some other exotic operating systems.
If you experience problems with the installation, we would suggest to run the experiments in a virtual environment, e.g., using Oracle's VirtualBox_.
On request, we will generate a virtual image with this package pre-installed.

Dataset
-------

This package does not include the original image and protocol files for the competition.
Please register on the competition website_ and download the UCCS dataset from there.

Please extract all zip files **into the same directory** (the .zip files contain the appropriate directory structure).
This includes all training_*.zip and validation_*.zip files, as well as the protocol.zip and possibly the SampleDataSet.zip.
The directory will be refered to as YOUR-DATA-DIRECTORY below.

Installation
------------

The installation of this package follows the Buildout_ structure.
After installing Bob_ and extracting this package, please run the following command lines to install this package:

$python bootstrap-buildout.py ...$ ./bin/buildout

...

The installation procedure automatically generates executable files inside the bin directory, which can be used to run the baseline algorithms or to evaluate the baselines (and your) algorithm.

Running the Baselines
---------------------

There are two scripts to run the baseline, one for each part.

Face Detection
~~~~~~~~~~~~~~

The first script is a face detection script, which will detect the faces in the validation (and test) set.
The baseline face detector simply uses Bob_'s built-in face detector bob.ip.facedetect_, which is neither optimized for blurry faces nor for profiles.
Hence, there are lots of misdetections (false negatives) and detected background areas (false positives).

You can call the face detector baseline script using:

$./bin/baseline_detector.py Please refer to ./bin/baseline_detector.py -h for possible options. Here is a subset of options that you might want/need to use/change: --data-directory: Specify the directory, where you have downloaded the UCCS dataset into --result-file: The file to write the detection results into; this will be in the required format --which-set: The set, for which the baseline detector should be run; validation by default --verbose: Increase the verbosity of the script; using --verbose --verbose or -vv is recommended; -vvv will write more information --debug: Run only over the specified number of images; for debug purposes only --display: Display the detected bounding boxes and the ground truth; for debug purposes only --parallel: Run in the given number of parallel processes; can speed up the processing tremendously On a machine with 32 cores, a good command line for the full baseline experiment would be:$ ./bin/baseline_detector.py --data-directory YOUR-DATA-DIRECTORY -vv --parallel 32

To run a small-scale experiment, i.e., to display the detected faces on 20 images, a good command line might be:

$./bin/baseline_detector.py --data-directory YOUR-DATA-DIRECTORY -vvv --display --debug 20 .. note:: The --display option requires Matplotlib_ to be installed and set up properly. Display does not work properly in parallel mode. By default, the face detection score file will be written to ./results/UCCS-detection-baseline.txt. Face Recognition ~~~~~~~~~~~~~~~~ For face recognition, we simply adopt a PCA+LDA pipeline on top of LBPHS features. The PCA+LDA projection matrix is estimated from the faces in the training set. For each person, the images of the training set build one class. Open-set recognition is performed by using all training faces of unknown identities in a separate class. First, the faces in the training images are re-detected, to assure that the bounding boxes of training and test images have similar content. Then, the faces are rescaled and cropped to a resolution of 64x80 pixels. Afterwards, LPBHS features are extracted from these crops, and a PCA+LDA projection matrix is computed. All training features are projected into the PCA+LDA subspace. For each identity (including the unknown identity -1), the average of the projected features is stored as a template. During testing, in each image all faces are detected, cropped, and LBPHS features are extracted. Those probe features are projected into the same PCA+LDA subspace, and compared to all templates using Euclidean distance. For each detected face, the 10 identities with the smallest distances are obtained -- if identity -1 is included, all less similar images are not considered anymore. These scores are written into the score file in the desired format. You can call the face recognition baseline script using:$ ./bin/baseline_recognizer.py

Please refer to ./bin/baseline_recognizer.py -h for possible options.
Here is a subset of options that you might want/need to use/change:

--data-directory: Specify the directory, where you have downloaded the UCCS dataset into

--result-file: The file to write the recognition results into; this will be in the required format

--verbose: Increase the verbosity of the script; using --verbose --verbose or -vv is recommended; -vvv will write more information

--temp-dir: Specify the directory, where temporary files are stored; these files will be computed only once and reloaded if present

--force: Ignore existing temporary files and always recompute everything

--debug: Run only over the specified number of identities; for debug purposes only; will modify file names of temporary files and result file

--display: Display the detected bounding boxes and the ground truth; for debug purposes only

--parallel: Run in the given number of parallel processes; can speed up the processing tremendously

On a machine with 32 cores, a good command line would be:

$./bin/baseline_recognizer.py --data-directory YOUR-DATA-DIRECTORY -vv --parallel 32 .. warning:: The processing, particularly the face detection, will take a long runtime. Even with 32 parallel processes, several hours of processing will be required. .. note:: During training image detection, you will observe several warnings of training faces not being detected. This is normal as the face detector was designed to detect frontal faces only. The processing will work without these faces being detected. By default, the face recognition score file will be written to ./results/UCCS-recognition-baseline.txt. Evaluation ---------- The provided evaluation scripts will be usable to evaluate the validation set only, not the test set. You can use the evaluation scripts for two purposes: 1. To plot the baseline results in comparison to your results. 2. To make sure that your score file is in the desired format. If you are unable to run the baseline experiments on your machine, we provide the score files for the validation set on the competition website_. Face Detection ~~~~~~~~~~~~~~ As the ground-truth is usually larger than the face, we do not punish bounding boxes that are smaller than the ground truth. Therefore, the union (the denominator) takes into account only one fourth of the ground truth boudning box -- or the intersection area, whichever is larger: .. math:: O(G,D) = \frac{|G \cap D|}{|G \cup D|} = \frac{G \cap D}{\max\{\frac{|G|}4, |G \cap D|\} + |D| - |G \cap D|} where :math:|\cdot| is the area operator. Hence, when the detected bounding box :math:D covers at least a fourth of the ground-truth bounding box :math:G and is entirely contained inside :math:G, an overlap of 1 is reached. The face detection is evaluated using the Free Receiver Operator Characteristic (FROC) curve, which plots the percentage of correctly detected faces over the total number of false positives (false alarms). This plot can be created using:$ ./bin/evaluate_detector.py

This script has several options, some of which need to be specified, see ./bin/evaluate_detector.py -h:

--data-directory: Specify the directory, where you have downloaded the UCCS dataset into

--result-files: A list of all files containing detection (or recognition) results

--labels: A list of labels for the algorithms; must be the same number and in the same order as --result-files

--froc-file: The name of the output .pdf file containing the FROC plot

--log-x: will plot the horizontal axis in logarithmic scale

--only-present: will ignore any file for which no detection exists (for debug purposes only, i.e., when detector ran with the --debug option)

--verbose: Increase the verbosity of the script; using --verbose --verbose or -vv is recommended

To plot the baseline FROC curve (which is shown on the competition website_), execute:

$./bin/evaluate_detector.py --data-directory YOUR-DATA-DIRECTORY --result-files results/UCCS-detection-baseline.txt --labels Baseline -vv .. note:: If you have run the face recognition baseline, you can also use the face recognition result file for plotting the FROC curve:$ ./bin/evaluate_detector.py --data-directory YOUR-DATA-DIRECTORY --result-files results/UCCS-recognition-baseline.txt --labels Baseline -vv

Face Recognition
~~~~~~~~~~~~~~~~

Open set face recognition is evaluated using the Detection and Identification Rate (DIR) curve, which plots the percentage of correctly detected and identified faces over the false alarm rate (FAR).
Based on various values of the FAR, several score thresholds are computed.
A face is said to be identified correctly if the recognition score is greater than the threshold and the correct identity has the highest recognition score for that face.
The number of correctly identified faces is computed, and divided by the total number of recognition scores greater than the threshold.
For more details, please refer to [1]_.

.. note::
By default only rank 1 recognition is performed, but the evaluation can be done using any rank up to 10 (the upper bound of allowed labels per face).
Providing more than one identity label per face will increase the number of false alarms, and may only have an impact on higher rank evaluations.

.. note::
Unknown identities or background regions labeled with label -1 or not labeled at all will be ignored (i.e., will not decrease performance).
Labeling an unknown identity or a background region with any other label than -1 will result in a false alarm -- only the maximum score per bounding box will be considered.

The DIR plot can be created using:

$./bin/evaluate_recognizer.py As usual, the script has several options, which are similar to ./bin/evaluate_detector.py above, see ./bin/evaluate_recognizer.py -h for a complete list: --data-directory: Specify the directory, where you have downloaded the UCCS dataset into --result-files: A list of all files containing recognition results --labels: A list of labels for the algorithms; must be the same number and in the same order as --result-files --dir-file: The name of the output .pdf file containing the DIR plot --log-x: will plot the horizontal axis in logarithmic scale --only-present: will ignore any file for which no detection exists (for debug purposes only, i.e., when recognizer ran with the --debug option) --verbose: Increase the verbosity of the script; using --verbose --verbose or -vv is recommended --rank: Use the given rank to plot the DIR curve To plot the baseline Rank 1 DIR curve (which is shown on the competition website_), execute:$ ./bin/evaluate_recognizer.py --data-directory YOUR-DATA-DIRECTORY --result-files results/UCCS-recognition-baseline.txt --labels Baseline -vv

Trouble Shooting
----------------

.. _bob: http://www.idiap.ch/software/bob
.. _oracle's virtualbox: https://www.virtualbox.org
.. _matplotlib: http://matplotlib.org
.. _buildout: http://www.buildout.org
.. _bob.ip.facedetect: http:/pythonhosted.org/bob.ip.facedetect
.. _competition website: http://vast.uccs.edu/Opensetface

.. [1] **P. Jonathon Phillips, Patrick Grother, and Ross Micheals** "Evaluation Methods in Face Recognition" in *Handbook of Face Recognition*, Second Edition, 2011.