Carl Database

Authors

Virginia Espinosa-Duró, Marcos Faundez-Zanuy and Jiří Mekyska

Access to the database

Please fill in a license agreement that can be downloaded in DOCX or PDF and send it to espinosa@tecnocampus.cat with CC to faundez@tecnocampus.cat and mekyska@feec.vutbr.cz. You will consequently get an access to the database.

Publications to be cited

V. Espinosa-Duró, M. Faundez-Zanuy and J. Mekyska, “A New Face Database Simultaneously Acquired in Visible, Near-Infrared and Thermal Spectrums”, Cognitive Computation, vol. 5, no. 1, pp. 119-135, 2013.

V. Espinosa-Duró, M. Faundez-Zanuy, J. Mekyska and E. Monte-Moreno, “A Criterion for Analysis of Different Sensor Combinations with an Application to Face Biometrics”, Cognitive Computation, vol. 2, no. 3, pp. 135-141, 2010.

Terms of use

This database is freely available. In case of any publications based on this database please cite the papers mentioned above.

Description

Visible and thermal images have been acquired using a thermographic camera TESTO 880-3, equipped with an uncooled detector with a spectral sensitivity range from 8 to 14 μm and provided with a germanium optical lens, and an approximate cost of 8.000 EUR. For the NIR a customized Logitech Quickcam messenger E2500 has been used, provided with a Silicon based CMOS image sensor with a sensibility to the overall visible spectrum and the half part of the NIR (until 1.000 nm approximately) with a cost of approx. 30 EUR. We have replaced the default optical filter of this camera by a couple of Kodak daylight filters for IR interspersed between optical and sensor. They both have similar spectrum responses and are coded as wratten filter 87 and 87C, respectively. In addition, we have used a special purpose printed circuit board (PCB) with a set of 16 infrared leds (IRED) with a range of emission from 820 to 1.000 nm in order to provide the required illumination.

The thermographic camera provides a resolution of 160×120 pixels for thermal images and 640×480 for visible images, while the webcam provides a still picture maximum resolution of 640×480 for near-infrared images and this has been the final resolution selected for our experiments.

A couple of halogen focus disposed 30 degrees away from the frontal direction and about 3 m away from the user, match the artificial light of the room. Note that all the tripods and structures have fixed markings on the ground.

Carl Database

Additionally, the distance between the face of the user and the tripod that holds the sensors is 135cm. This is in order to minimize the inherent parallax error in short distances between visible and thermal images of thermal imager and also for obtaining a similar field of view between these images and the near infrared images acquired with the customized webcam.

We have designed a background screen using a special stand kit which supports a roll of matt black paper. It is important to point out that this matt black background is mandatory behind the user in order to avoid undesirable thermal reflections from the operator, due to its well-known extra low albedo. This smooth background also facilitates the segmentation of the visible and NIR images.
In each recording session the images have been acquired under three different illumination conditions:

  • Natural illumination (NA): windows are open and sunlight enters the room. Obviously this illumination is not constant along days (due to weather conditions) and it also varies in function of the different hours of the day.
  • Infrared Illumination (IR): printed circuit board around the webcam is turned on and the remaining sources of light are disconnected. A graphic user interface has been developed in order to set properly the IRED’s intensity level and to set the image involved parameters (exposure, gamma and brightness). Additionally, it is also possible to manually fully optimize them.
  • Artificial Illumination (AR): The provided equipment used for illumination is the following: A set of 9 cool white fluorescents uniformly distributed in order to produce the base illumination of the scene has been used. A second pair of IANIRO Lilliput lights fitting 650W-3,400K tungsten halogen lamps have also been used in order to fill and smooth the well-known discontinuous fluorescent spectral emission and to provide and additional IR portion of light.

At the beginning, high pair of power focus produced important dark shadows over the users’ face. In order to solve this drawback, we had finally used a LEE 3ND 209 Filter to minimize the referred effect. This neutral density gel reduces light without affecting colour balance.

Each user has been recorded in four different acquisition sessions performed between November of 2009 and January 2010. In this sense, distinctive changes in the haircut and/or facial hair of some subjects may be appreciated. The acquisitions have been done in the whole day from 9 AM to 5 PM, because it was getting dark after 5 PM. The average time required for the full acquisition process of a skilled user has been 10 minutes, being 15 minutes for a non-skilled one. The whole set of users were acquired in two days per session.

In each illumination condition five different frontal snapshots are acquired. During the acquisition process, the user is required to look straight at the same place. No keeping neutral facial expression is required. Thus, a different facial expressions have been collected (smiling/non smiling, open-closed and blinking eyes…etc). Due to glasses exhibit a fully different behavior as function of the spectrum, being transparent from the VIS to the NIR spectrum and fully opaque beyond 3 μm approximately, people wearing glasses were asked to remove them before acquisition. No any other physical restriction has been taken into account in order to acquire a face image.

In order to reduce the correlation between consecutive acquisitions of the same session, between a couple of snapshots the user is asked to stand up make a loop to the room, including one step which corresponds to the portion of the room close to the blackboard, and sit down again. It is worth to mention that thermal camera was able to detect a temperature increase due to this additional physical exercise.

Final database consists of 41 people (32 males, 9 females). Each individual contributed in four acquisition sessions and provided five different snapshots in three different illumination conditions and under three image sensors. This implies a total of: 41x4x5x3x3= 7.380 images.

The images in NIR spectrum are stored in lossless *.bmp files. The images from thermal camera were firstly stored to *.bmt format provided by TESTO company. This file includes VIS image, temperature matrix and metadata describing for example the outside humidity, temperature range etc. This file was processed and the image in VIS spectrum was extracted. The temperature matrix was stored to MATLAB *.mat file and also transformed to grayscale image and stored to *.bmp format.