1996 FERET Test Results

In September 1996, we participated in the final round of FERET tests administered by the US Army Research Laboratory which consisted of a large gallery test containing nearly 3,800 frontal images. Our system was found to be the top competitor (by a typical margin of 10% to the next best competitor). The details of this competition and an analysis of the test results are available in this PostScript preprint by Jonathan Phillips.

Our new probabilistic matching algorithm (a simpler version of the system described in Tech-Report #393) obtained a 95% rank-1 recognition accuracy on pairs of frontal images (FA vs. FB) and a 60% rank-1 recognition accuracy on "duplicate" images (images of the individual taken weeks or months later). These recognition rates were the highest obtained by any of the other 5 competing algorithms in the final testing round. The plots below demonstrate the performance of our new Bayesian algorithm as compared to our previous method which used a standard eigenface nearest-neighbor classifier. Performace curves for the other participants are also shown (see legend in the graphs).

NOTE:

To obtain a copy of the official government report on the FERET program and test results, contact Jonathan Phillips.

FA vs. FB Recognition Performance Results of large gallery test from September 1996. FA/FB are two views taken at the same time (with different expressions). The top curve is our new Bayesian similarity matching algorithm. The older version of our system, with standard nearest-neighbor eigenface matching algorithm is also shown, along with the other competitors. PostScript image

Duplicate Recognition Performance Results of large gallery test from September 1996. Duplicates are images of the same person taken weeks/months apart. The top curve is our new Bayesian similarity matching algorithm. The older version with standard nearest-neighbor eigenface matching algorithm is also shown, along with the other competitors. PostScript image

Older Duplicate Recognition Performance Results of large gallery test from September 1996. In this graph performance on duplicate images taken at least a year apart is shown. The top curve is our new Bayesian similarity matching algorithm. The older version with standard nearest-neighbor eigenface matching algorithm is also shown, along with the other competitors. PostScript image