Big IR dataset compiled for training face biometric systems
The U.S. Army has created a public, 500,000-image dataset that officers hope will lead to better low-light facial recognition systems.
Heat-sensitive cameras can pick up blobs representing people’s faces captured in the dark, but the blobs typically are useless for confidently identifying someone. Biometrics are identifying people through COVID masks, and yet darkness remains a frustrating factor for efforts to recognize faces outside lab settings.
In a paper yet to be peer reviewed, a large team of Army, university and private sector researchers created the new time synchronized Army Research Lab Visible-Thermal Face database from images of 395 subjects.
Authors of the paper say they believe they have collected the world’s largest biometric dataset of paired visible and thermal face images.
A long wave infrared camera was mounted alongside multiple visible-light cameras that created 3D images. Each subject was captured in a straight-on baseline image, one showing an expression and a third that was posed so that he or she was not looking directly into a lens. Subjects who wear eyeglasses were allowed to wear them in a fourth sequence.
Joining the Army’s Devcom research lab in the work were the West Virginia University, Booz Allen Hamilton, Johns Hopkins University and the University of Nebraska-Lincoln.