FaceVACS Performance

Various tests by independent organizations have demonstrated that
Cognitec's face recognition technology is industry-leading in terms of
recognition accuracy and speed.

FRVT 2002: At the Face Recognition Vendor Test 2002, conducted by the
National Institute of Standards and Technology (NIST), Cognitec's face
recognition software exhibited the best performance among all
participants for large-scale verification, identification, as well as
watch list scenarios.

FRVT 2006: The Face Recognition Vendor Test 2006, again conducted by
NIST, confirmed the outstanding performance of Cognitec's
technology. Of 22 companies and research institutes, only Cognitec and
10 other participants could complete the large-scale tests, and only
Cognitec and one other company could deliver measurements in four
different test scenarios for still images and 2D/3D data.

The results also demonstrated that Cognitec exceeded the goal
specified by the preceding so-called Face Recognition Grand Challenge,
or FRGC, which was to prove increased recognition performance by an
order of magnitude as compared to results measured during FRVT
2002. Cognitec again delivered top performance in the majority of the
tests, especially with regard to common real-world scenarios (low
resolution imagery).

MBE 2010: The latest NIST test, conducted in 2010, consists of three
parts, of which only the results of the first part, MBE-STILL, have
been published so far. The purpose of MBE-STILL has been to measure
and compare the recognition performance of face recognition algorithms
on still images.

Seven companies and three research institutes participated in the
MBE-STILL evaluation. Cognitec again delivered excellent performance
in all the tests, especially in the more challenging tests that
reflect common real-world scenarios like searching through very large
photo databases of passport or drivers' license authorities. In the
open-set identification tests, which is the practically most
relevant test for those challenging applications, Cognitec's algorithm
had the best performance at the lowest false alarm levels, which
corresponds to the appropriate operational setting in case of a
multi-million gallery and limited resources for visual inspection.

Disclaimer: FRVT and MBE results do not constitute endorsement of any
particular system by the U.S. Government. Complete test results can be
downloaded at face.nist.gov.

We are currently in the process of productizing the B5T8 algorithm,
which is an improved version of the algorithm submitted to the
MBE-STILL test. Until that process is finished, the performance
achievable with our product software is that of the algorithm B4T8.
Performance results in the Grayscale FERET 'Duplicate I' test, which
uses images from the public FERET database, can be downloaded here.

FaceVACS-Performance

The 'Duplicate I' test involves the following subsets of the Grayscale
FERET database: a gallery including 1196 images of 1183 persons and a
probe set of 722 images of 242 persons. The lower curve is based on
the FaceVACS engine used during FRVT2006, the upper curve uses the
current FaceVACS engine version B4T8.

Neither of Cognitec's algorithms is specifically optimized or trained
on databases used for tests, like the FERET database. Training and
optimization is only done on internal proprietary databases which do
not contain data from test databases. Consequently, Cognitec's test
results can be generalized to similar unknown set of data.

MBGC 2009

MBGC is a project conducted by the U.S. National Institute of Standards and Technology (NIST) with the purpose to foster progress in face and iris recognition technology. The Multiple Biometric Grand Challenge is not considered an independent evaluation test as the MBGC participants report their results to NIST themselves, a circumstance not allowing a practically relevant comparison of the capabilities of different face recognition technologies.

For the MBGC, NIST defines several so-called challenge problems (involving comparing face images and videos taken under various, more or less difficult conditions), and prospective participants are invited to apply their technology to those problems and to submit their results, in terms of comparison scores. All image data used in the challenge problems are public, enabling the participants to tune their algorithms to the image data.