IARPA, NIST hosting face recognition challenge to improve accuracy
The Intelligence Advanced Research Projects Activity (IARPA), within the Office of the Director of National Intelligence (ODNI), has launched the Face Recognition Prize Challenge (FRPC) with a total of $50,000 in cash prizes.
The challenge, which is organized in partnership with National Institute of Standards and Technology, is designed to improve core face recognition accuracy and expand the breadth of capture conditions and environments suitable for successful face recognition.
The Challenge if comprised of two components. The face identification component involves executing one-to-many search to return the correct entry from a gallery, if any.
The second part deals with face verification, which requires the submitted algorithm to match two faces of the same individual while correctly rejecting faces of different individuals. Both tasks involve “non-cooperative” images where subjects were either unaware of the camera or did not pose for the camera.
IARPA is inviting all academic and industry developers of automated face recognition algorithms to participate by submitting pre-compiled software libraries to NIST, who is the designated test laboratory for the FRPC.
At NIST the algorithms will be run on sequestered images, meaning that FRPC is not an “open-book’’ test so no test or training images will be made available to developers.
A $25,000 prize will be awarded to the applicant with the highest one-to-many identification accuracy, $20,000 for the highest one-to-one verification accuracy, and $5,000 for the highest one-to-many identification speed.
The contest is being assessed by a panel of judges, which includes Chris Boehnen Ph.D., senior program manager at IARPA; Patricia Wolfhope, program manager at DHS S&T; and Patrick Grother, staff scientist at NIST.
The primary prize winner will be declared by considering measurements of identification accuracy. This will be stated as the False Negative Identification Rate (FNIR) measured at the lowest scalar threshold that gives a fixed False Positive Identification Rate (FPIR) no higher than 10^-3.
FNIR, which is the proportion of mate searches for which a correct mate is not returned above a threshold, will be measured over many mate searches. Mate searches are those for which the person in the search image has a face image in the enrolled dataset.
FPIR, which is the proportion of non-mate searches that yield one or more non-mates at or above threshold, will measured over many non-mate searches. Non-mate searches are those which the person in the search image does not have a face image in the enrolled dataset.
The conduct of both mate and non-mate searches defines an open-set issue. If two participants happen to share the same identification accuracy or verification accuracy, the algorithm with the lowest median search duration or lowest median template generation duration, respectively, will be declared the primary prize winner.
Interested developers can visit the Challenge website where they will be redirected to the National Institute of Standards and Technology (NIST) FRPC support website. Registration officially closes on June 15.