Pgu-Face: A dataset of partially covered facial images

In this article we introduce a human face image dataset. Images were taken in close to real-world conditions using several cameras, often mobile phone׳s cameras. The dataset contains 224 subjects imaged under four different figures (a nearly clean-shaven countenance, a nearly clean-shaven countenance with sunglasses, an unshaven or stubble face countenance, an unshaven or stubble face countenance with sunglasses) in up to two recording sessions. Existence of partially covered face images in this dataset could reveal the robustness and efficiency of several facial image processing algorithms. In this work we present the dataset and explain the recording method.


a b s t r a c t
In this article we introduce a human face image dataset. Images were taken in close to real-world conditions using several cameras, often mobile phone's cameras. The dataset contains 224 subjects imaged under four different figures (a nearly clean-shaven countenance, a nearly clean-shaven countenance with sunglasses, an unshaven or stubble face countenance, an unshaven or stubble face countenance with sunglasses) in up to two recording sessions. Existence of partially covered face images in this dataset could reveal the robustness and efficiency of several facial image processing algorithms. In this work we present the dataset and explain the recording method.
& Upper and lower face occlusion is one the main specific of this new dataset. Some facial image algorithms work well on lower face occlusion while not well for upper occlusion. Face upper regions contain the most human identity information, then upper occlusion covers these information [1]. This new dataset covers upper occlusion in addition to lower occlusion, and therefore can be used to evaluate facial image processing performance on the both regions.
Age of each subject was provided in a text file with subject images, therefore it suitable for showing robustness of age estimation systems against occlusion problem.

Data
The Pgu-Face dataset contains 896 images from 224 different subjects. All of the subjects were Iranian men and most of them live in tropical regions of the southwest of Iran. The age range of the subjects was 16-82 years with average 27.89 years.

Data collection method
We captured four face images from each subject in two sessions with different figures. During each session, we recorded two neutral frontal images. In the first session, for each person, two images were recorded in such a way that in an image the subject's face was not covered by any cover such as a beard, mustache, glasses, etc., and was in a nearly clean-shaven countenance. In another image, the subject wore sunglasses. In the next session, for each subject, two images were recorded, including an image of the subject's face covered only with a beard and mustache, and then in another image the Table 1 Four different subject's countenance images recorded in two sessions. (two sessions were separated with 6 day minimum time between them).

Session 1 Session 2
First image In a nearly clean-shaven countenance In a stubble or unshaven face countenance Second image In a nearly clean-shaven countenance with sunglasses In a stubble or unshaven face countenance with sunglasses subject wore sunglasses. The minimum period between these two sessions was six days. Images were taken from about fifty to one hundred centimeters and subject's faces were in neutral expression without any specific angle to the camera. The images were taken in different locations, where the locations often were roofed. No specific camera stands were applied to position cameras. The light at each location was natural and most of the images were captured at night. For every subject, as described before, images were denominated as img01 to img04, respectively. Table 1 lists four different subject's figures. Fig. 1 shows all images from two subjects for two sessions.

Resolution of images
The resolution of the used cameras varies in range of 2-26 mega pixels. Therefore, the images dimensions varied over the used cameras. Most of the utilized cameras were commercial and mobile phone cameras. Mobile phone cameras against professional cameras have a lower quality and hence were suitable for our purpose, although they may have a better performance versus surveillance cameras. No necessary settings for all cameras were applied and all of the recorded images were captured in conventional conditions. The resolution of used cameras in addition to the number of images in each resolution were shown in Table 2. Based on achieved experimental results, it can  deduced that the existence of facial occlusion, such as glasses, beards and mustache on the face, decrease probability of recognition and reliability of system. Hence, Pgu-Face dataset can be used to challenge recently presented facial image processing algorithms.

Transparency document. Supplementary material
Supplementary data associated with this article can be found in the online version at http://dx.doi. org/10.1016/j.dib.2016.09.002.