Biometric recognition on the CPqD Biometric Database (BioCPqD Phase 1)

1. All publications that report on research that use the Corpus on the BEAT platform will acknowledge the BioCPqD database and the BEAT platform by referring to the following publication:

   @article{Violato_CPqD_2013,
     author = {R. P. V. Violato and M. Uliani Neto and F. O. Simoes and T. F. Pereira and M. A. Angeloni},
     title = {BioCPqD: uma base de dados biometricos com amostras de face e voz de individuos brasileiros},
     year = {2013},
     journal = {Cadernos CPqD Tecnologia, Campinas, Brazil},
     volume = {9},
     pages = {7--18},
     publisher = {CPqD},
     url = {http://www.cpqd.com.br/cadernosdetecnologia/Vol9_N2_jul_dez_2013/artigo1.html},
   }

2. Bob as the core framework used to run the experiments:

   @inproceedings{Anjos_ACMMM_2012,
     author = {A. Anjos and L. El Shafey and R. Wallace and M. G\"unther and C. McCool and S. Marcel},
     title = {Bob: a free signal processing and machine learning toolbox for researchers},
     year = {2012},
     month = oct,
     booktitle = {20th ACM Conference on Multimedia Systems (ACMMM), Nara, Japan},
     publisher = {ACM Press},
     url = {http://publications.idiap.ch/downloads/papers/2012/Anjos_Bob_ACMMM12.pdf},
   }

Overview

This database was designed to provide data that was recorded in a natural way, using various devices in different environments. Hence, algorithms that perform well on this database are expected to be suitable for other real-world applications that do not require a predefined audio/video recording setup.

Database participants were selected among employees of CPqD Foundation who volunteered to make recordings. A unique ID was assigned for each participant, composed by a prefix (M for male and F for female) followed by a 4-digit number (odd for males and even for females). Each participant recorded up to five sessions, with a time lapse of at least 10 days between sessions.

Sessions consisted of 27 recorded sentences, whose content was specified in a script. Each sentence was recorded on three different devices types:

• Laptops (audio and video content);
• Smartphones (audio and video content);
• Phone calls (only audio).

For each device type, a set of devices was used, as specified below:

• Laptops: - Compaq 510 with embedded mic and camera; - Toshiba with USB Logitech QuickCam Pro 9000 webcam; - DELL Latitude embedded mic and camera.
• Smartphones: - Samsung Galaxy S II; - Apple iPhone 4; - Apple iPhone 4.
• Phone calls: - landline phone call; - personal mobile phone call.

Recordings were made in three environments with different characteristics: garden, restaurant (public indoor) and office. The idea behind this strategy was to exploit the influence of environmental noise in audio recordings and the effect of illumination and background conditions in the video recordings. Since the database includes recordings captured on different devices of different types and in different environments, it allows a large number of experimental setups.

Content

The data collection followed a simple recording protocol that was replicated for all sessions. For each session there was a corresponding script describing the whole content to be recorded, as follows:

Text reading:

• a pre-defined text (extracted from the database's consent form);
• four phonetically rich sentences (randomly selected among 562 options);
• passphrase: three repetitions a single sentence (the same sentence for all participants in all sessions).

Spontaneous speech:

• answers for generic questions (all participants answered all 15 questions selected form a fixed set, distributed along the 5 sessions in random order);
• a fake name;
• a fake address;
• a fake birthday date;
• a fake ID number;
• a fake phone number;
• two command words (all participants spoke 10 words along the 5 sessions in random order).

Numbers, digits, time values and alphanumeric strings:

• a monetary amount between 10 and 10 000, randomly generated;
• a number between 10 and 1000, randomly generated;
• a number between 1000 and 10 million, randomly generated;
• three repetitions of a random digit sequence (first one read in a slow pace and others naturally read);
• a fake credit card number;
• an alphanumeric string composed of 6 characters, randomly generated;
• a time value, selected among a predefined set with 181 samples, equally distributed among participants.

It is important to note that all content was recorded in Brazilian Portuguese language.

BioCPqD Phase I database provides unbiased biometric verification protocols, one for male and one for female participants, based on the MOBIO database protocols. These protocols partition the database in three different groups: - a Training set: used to train the parameters of algorithm to be tested, e.g., to create the projection matrix, Universal Background Models, etc.; - a Development set: used to evaluate hyper-parameters of the tested algorithms; - a Test set: used to evaluate the generalization performance of the tested algorithms with previously unseen data.

Both development and test sets are further split into an enrollment subset (used to enroll participants' models), and a probe set (whose files will be tested against all participants' models).