Face Recognition Vendor Test 2002 Performance Metrics

Audio-Visual Based Person Authentication, in June 2003. The paper details the metrics used to quantify the performance of the face recognition systems tested in FRVT 2002. The methods are suited to any recognition evaluation, online or offline, technology or scenario, for which complete similarity scores are archived. The paper shows that the open-set identification problem known as the watch list task is the general case: it requires systems to perform 1:N recognition with concurrent possibilities of false acceptance and rejection. Two special cases are demonstrated: 1:1 Verification is simply the watch list task with N=1; and closed-set identification is that with no false accept rate. The paper also presents the computation of standard error ellipses used to show the effect of population variation on false accept and false reject rates