UPM-GTI-Face Dataset

Description

The UPM-GTI-Face dataset has been created with the purpose of studying the impact of distance and the use of face masks on the results of face detection and recognition strategies.

The UPM-GTI-Face is composed of images from 11 different subjects (8 men and 3 women) captured in 4K, under 2 scenarios (indoors and outdoors), and 2 face mask conditions (subjects with and without face masks). For each combination of subject, scenario, and face mask condition, the database includes a mugshot image at a distance of 1 meter, and 10 probe images acquired from 3 to 30 meters (at 3-meter intervals). As a result, UPM-GTI-Face contains 484 face images (22 indoor gallery images, 22 outdoor gallery images, 220 indoor probes and 220 outdoor probes). Figure 1 shows a sample of the images for one of the subjects in the dataset.

Figure 1. UPM-GTI-Face dataset sample images for one of the subjects in the dataset. 11 different subjects were captured under different scenarios and conditions. Scenarios: indoor and outdoor. Conditions: mask and no mask. For every combination of scenario, condition, and subject in the dataset, there is 1 close-up shot for gallery, and 10 probe images that correspond to each distance mark.

In an attempt to imitate surveillance scenarios, images were acquired using a Logitech BRIO Webcam camera placed on a tripod, at a height of 2.1 meters (see Fig. 2). People were asked to walk towards the camera, starting from a distance of 30 meters, stopping for a couple of seconds at several floor marks indicating intermediate distances, without any further cooperation from them (e.g., pose or expression). A continuous video was recorded for these sequences, while the supervisor voiced aloud the moments at which users stopped at the distance marks. This facilitated the process of extracting the frames corresponding to each distance using audio as a cue.

Figure 2. UPM-GTI-Face dataset capturing setup for both indoor and outdoor scenarios.

The UPM-GTI-Face dataset is structured as described in Table 1. The dataset contains two main folders, one for the indoor scenario, and another one for the outdoor scenario. Each scenario is divided by face mask condition into two folders, one for masked subjects, and another one for non-masked subjects, both of which are also divided into gallery and probe folders. All images from the 11 different subjects can be located in their respective folders.

 Table 1. UPM-GTI-Face dataset structure.

Download

Click here to download the UPM-GTI-Face dataset.

Citation

M. Rodrigo, E. Gonzalez-Sosa, C. Cuevas, and N. Garcia, “"UPM-GTI-Face: A dataset for the evaluation of the impact of distance and masks in face detection and recognition systems”, under review for AVSS 2022.