Other Free Encyclopedias » Online Encyclopedia » Encyclopedia - Featured Articles » Contributed Topics from F-J

Face Recognition Evaluation and Pilots

images performance database frr

Definition: Standard evaluation protocols are used to evaluate various face recognition algorithms and examine their limitations.

The growing number of face recognition (FR) methods has imposed the development of standard evaluation protocols and the creation of large evaluation databases. Based on that, a series of public tests evaluating face recognition algorithms and examining their limitations have been performed. The most recent one is the Facial Recognition Vendor Test of 2002 (FRVT2002), in which several state-of the art commercial FR systems were tested using a database of 121,589 images of 37,437 Mexican VISA applicants collected by the US Department of State. The verification performance was measured using the false acceptance (FAR) and false rejection (FRR) rates. FAR is defined as the percentage of instances that a non-authorized individual is falsely accepted by the system, while FRR is defined as the percentage of instances an authorized individual is falsely rejected by the system. In an indoor environment the best systems had an 18% FRR for a FAR of 0.1% (user unfriendly, high-security application), a 10% FRR for a FAR of 1% and a 4% FRR for a FAR of 10% (user friendly application). For test images taken outdoors the FRR increased by 42% (from 4% to 46%) at FAR=10%, showing that FR under illumination variations is still an open issue.

The identification performance was measured in terms of the percentage of images that are correctly identified as the best match. The best performance was 85% on a database of 800 people, 83% on a database of 1,600 and 73% on a database of 37,437, and showed that identification performance decreases linearly in the logarithm of the database size. The identification rates for males were 6%-9% higher than for females. Moreover, it was found that for every 10 years decrease in age the identification performance decreased by approximately 5% (old people are easier to recognize). A reduction of 5% per year of elapsed time between enrolled and new images was also observed. Finally, it was shown that the use of deformable face models could significantly improve recognition under varying pose: for non-frontal images the identification performance of the best system was as low as 26%. When such models were employed for pose compensation, the performance increased impressively to 84% .

During the last years, FR systems have been employed by government agencies in several pilot applications. For example, a network of 250 CCTV cameras for surveillance was installed in the London Borough of Newham and has succeeded to reduce local crime. The cameras scan the faces within their view and compare them against a database of criminals. If a match is found, an alert is sent to the Police; otherwise the images are automatically discarded to ensure privacy. A FR system performing a face-to-passport check has been installed in Sydney’s Airport to verify the identity of passengers and automate airport controls. If there is a successful match, immigration and customs checks are made by the computer to ensure that border crossing is legitimate, and if so, an electronic gate opens for the passenger to pass. Another example of a FR application with an extremely important social impact is the ChildBase System, which organizes and classifies millions of Internet images of sexually abused children. FR technology is used to identify victims of abuse and their offenders by crosschecking un-indexed images with the world’s largest database of child abuse imagery, in order to quickly determine whether a seized computer contains images of known or new abuse victims.

Facial Angle [next] [back] Face Recognition, 3D

User Comments

Your email address will be altered so spam harvesting bots can't read it easily.
Hide my email completely instead?

Cancel or