xbob.paper.tpami2013 0.2.0a4

Getting the data

You first need to buy and download the Multi-PIE database:

http://multipie.org/

and to download the annotations available here:

http://www.idiap.ch/resource/biometric/

Feature extraction

The following command will extract LBP histograms features. You should set the paths to the data according to your own environment:

$ ./bin/lbph_features.py --image-dir /PATH/TO/MULTIPIE/IMAGES --annotation-dir /PATH/TO/MULTIPIE/ANNOTATIONS --output-dir /PATH/TO/MULTIPIE/OUTPUT_DIR/

Dimensionality reduction

Once the features has been extracted, they are projected into a lower dimensional subspace using Principal Component Analysis (PCA):

$ ./bin/pca_features.py --output-dir /PATH/TO/MULTIPIE/OUTPUT_DIR/

Proposed system: PLDA modeling and scoring

PLDA is then applied on the dimensionality reduced features.

This involves three different steps:

1. Training

2. Model enrollment

3. Scoring

The following command will perform all these steps:

$ ./bin/toolchain_plda.py --output-dir /PATH/TO/MULTIPIE/OUTPUT_DIR/

Then, the HTER on the evaluation set can be obtained using the evaluation script from the bob library as follows:

$ ./bin/bob_compute_perf.py -d /PATH/TO/MULTIPIE/OUTPUT_DIR/U/plda/scores/scores-dev -t /PATH/TO/MULTIPIE/OUTPUT_DIR/U/plda/scores/scores-eval -x

The HTER on the evaluation set, when using the the EER on the development set as the criterium for the threshold, corresponds to the PLDA value reported on Table 3 of the article mentioned above.

If you want to reproduce the Figure 2 of the article, you can run the following commands (instead of the previous one):

$ ./bin/experiment_plda_subworld.py --output-dir /PATH/TO/MULTIPIE/OUTPUT_DIR/
$ ./bin/plot_figure2.py --output-dir /PATH/TO/MULTIPIE/OUTPUT_DIR/

Then, the value of the HTER on Table 3 of the article (for the PLDA system) corresponds to the one, where the full training set is used, and might similarly be obtained as follows:

$ ./bin/bob_compute_perf.py -d /PATH/TO/MULTIPIE/OUTPUT_DIR/U/plda_subworld_76/scores/scores-dev -t /PATH/TO/MULTIPIE/OUTPUT_DIR/U/plda_subworld_76/scores/scores-eval -x

The previous commands will run the PLDA toolchain several times for a varying number of training samples. Please note, that this will require a lot of time to complete (one to two days on a recent workstation such as one with an Intel Core i7 CPU).

Baseline 1: PCA on the LBP histograms

The LBP histogram features were used in combination with the PCA classification technique (commonly called Eigenfaces in the face recognition litterature).

This involves three different steps:

1. PCA subspace training

2. Model enrollment

3. Scoring (with an Euclidean distance)

The following command will perform all these steps:

$ ./bin/toolchain_pca.py --n-outputs 2048 --output-dir /PATH/TO/MULTIPIE/OUTPUT_DIR/

Then, the HTER on the evaluation set can be obtained using the evaluation script from the bob library as follows:

$ ./bin/bob_compute_perf.py -d /PATH/TO/MULTIPIE/OUTPUT_DIR/U/lda/scores/scores-dev -t /PATH/TO/MULTIPIE/OUTPUT_DIR/U/lda/scores/scores-eval -x

This value corresponds to the one of the PCA baseline reported on Table 3 of the article (Once more, be aware of differences due to the implementation changes in the feature extraction process and algorithm parameters that have not been kept). These results are obtained for a PCA subspace of rank 2048, which was found to be the optimal PCA subspace size, when we tuned this parameter using the LBPH features.

Baseline 2: LDA on the PCA projected LBP histograms

The PCA projected LBP histogram features considered for the PLDA system were also used in combination with the Fisher’s LDA classification technique (commonly called Fisherfaces in the face recognition litterature).

This involves three different steps:

1. LDA subspace training

2. Model enrollment

3. Scoring (with an Euclidean distance)

The following command will perform all these steps:

$ ./bin/toolchain_lda.py --output-dir /PATH/TO/MULTIPIE/OUTPUT_DIR/

Then, the HTER on the evaluation set can be obtained using the evaluation script from the bob library as follows:

$ ./bin/bob_compute_perf.py -d /PATH/TO/MULTIPIE/OUTPUT_DIR/U/lda/scores/scores-dev -t /PATH/TO/MULTIPIE/OUTPUT_DIR/U/lda/scores/scores-eval -x

This value corresponds to the one of the LDA baseline reported on Table 3 of the PLDA article (Once more, be aware of slight differences due to the implementation changes in the feature extraction process). These results are obtained for a LDA subspace of rank 64, which was found to be the optimal LDA subspace size, when we tuned this parameter using the initial LBPH features.

Baseline 3: LBP histogram classification with Chi square scoring

The LBP histogram features might be used in combination with a distance such as the Chi Square distance, to obtain a face recognition system.

This involves two different steps:

1. Model enrollment

2. Scoring (with a chi square distance)

The following command will perform all these steps:

$ ./bin/toolchain_lbph.py --output-dir /PATH/TO/MULTIPIE/OUTPUT_DIR/

Then, the HTER on the evaluation set can be obtained using the evaluation script from the bob library as follows:

$ ./bin/bob_compute_perf.py -d /PATH/TO/MULTIPIE/OUTPUT_DIR/U/lbph_chisquare/scores/scores-dev -t /PATH/TO/MULTIPIE/OUTPUT_DIR/U/lbph_chisquare/scores/scores-eval -x

This value corresponds to the one of the LBP histogram (chi square) baseline reported on Table 3 of article (Once more, be aware of slight differences due to the implementation changes on the feature extraction process).

Summarizing the results as in Table 3

If you successfully run all the previous experiments, you could get a summary of the performances, as in Table 3, by running the following command:

$ ./bin/plot_table3.py --output-dir /PATH/TO/MULTIPIE/OUTPUT_DIR/