COMPACT Database
Biometric Dataset of Face Images Acquired in Uncontrolled Indoor Environment

Description

Detailed description of the COMPACT database is given in [1]. This page provides the most relevant information.

The COMPACT database aims at fostering the development of biometric recognition systems that work indoors and in unconstrained conditions. It collected with the dedicated acquisition system constructed by the authors. It is based on two wide-field of View (WFOV) cameras and one narrow-field of View (NFOV) camera. The WFOV cameras operate in the visible light spectrum, observe the entire scene and locate potential subjects that are to be identified. They form a stereo pair so they need to be precisely synchronized with each other. Once the system detects that the subject’s face is visible and its distance to the system is sufficient to perform the recognition, the NFOV camera is directed to capture high resolution biometric images. In order to provide stable lighting conditions, the NFOV images are acquired in near-infrared light. The described concept is presented in Fig. 1.

compact db concept

Fig. 1. Design concept of the COMPACT recognition gate.

The COMPACT dataset was collected in an indoor laboratory environment. The acquisition consists of three parts: soft biometrics survey, multi-pose registration images and probe data. Registration and probe images were taken on different days to ensure reliable comparisons. The most important features of the COMPACT database are summarized in Table I.

Spectrum Near Infrared
Number of subjects 108
Number of registration images 12312
Number of probe images 31078
Avg. no. of probe images per subject 287
Image size

2560x2048 px (registration)
3520x2200 px (probe)

The registration data consists of face images acquired with the use of the dedicated rotating platform. The images were collected in the range of -45° to 45° degrees , every 5° , and within three vertical poses: looking ahead, looking up and looking down. It means that a single set consists of 19 ∗ 3 = 57 images. For each person, two consecutive sessions were acquired. The reason for duplicating the data was that the likelihood of acquiring degraded image of a selected pose (e.g. person blinks or twitches during frame acquisition) is significantly smaller. Sample images included in the registration data are presented in Fig. 2.

compact db registration samples

Fig. 2. Sample images of registration data acquired using the dedicated rotating platform.

The probe data is divided into two groups: high resolution face images captured by the NFOV camera and tracking sequences recorded by the WFOV vision system. A single recording contains images of one person walking through the recognition gate. All data was acquired in a fully automated way using the constructed acquisition system. In order to have a proper gradation of difficulty in the acquired data, the authors proposed 4 scenarios for the recordings. The first two scenarios are controlled, the other two are fully uncontrolled. People, who have glasses, were asked to wear them only in the last scenario. Each subject performed each scenario twice. In total, 8 recordings for each person were collected. Fig 3. presents sample NFOV and WFOV image sets. The following scenarios were proposed:

  • Scenario 1: A person walks through the recognition gate looking straight ahead. It allows to capture clear frontal face images. Such data can be used e.g. for the development process or to verify algorithms implementation
  • Scenario 2: A person walks through the recognition gate turning his/her head to the left (first recording) or right (second recording) without changing pitch and roll angles. The head is always turned away from the camera to make the images more difficult. This data can be used to study the impact of pose variations on recognition performance or head pose modelling
  • Scenario 3: A person walks through the recognition gate looking around the scene. This involves turning the head to the left, right, top and bottom. These images represent the real world uncontrolled data, including a mix of frontal and pose images
  • Scenario 4: A person walks through the recognition gate looking around the scene, as in the third scenario. However, during these recordings people wiped their eyes, corrected hairstyles or scratched their cheeks. This allowed for the acquisition of real world unconstrained images with occlusions

compact db probe samples

Fig. 3. Sample images from the probe data (WFOV_LEFT, WFOV_RIGHT, NFOV).

How to obtain

The database is publicly available to research and educational institutions.

To obtain the database fill in the release agreement and send it to the DMCS Biometric Laboratory.

References

[1] M. Włodarczyk, D. Kacperski, W. Sankowski, K. Grabowski, "COMPACT: Biometric Dataset of Face Images Acquired in Uncontrolled Indoor Enviroment", 2017