FpVTE was conducted primarily to assess the current capabilities of fingerprint identification technologies in one-to-many mode using operational datasets containing several million subjects. This was the first large-scale one-to-many fingerprint evaluation conducted at NIST in its history. The main feature of the tests is the large scale database sand various data types. The current FpVTE used a testing model closer to real one-to-many identification systems. The number of subjects used was also significantly higher, about 10 million subjects in the testing data sets.
There were three classes of participation that examined one-to-many identification using various finger combinations from single finger up to ten fingers.
Class A used index fingers capture data and evaluated index fingers identification with out segmentation on database of 1.6 million records.
Class B used identification flat captures (4-4-2; left slap, right slap, and two thumbs simultaneously) and evaluated ten-finger, eight-finger and four-finger identification with segmentation on database of 3 million records.
Class C used rolled and plain impression (4-4-1-1; left slap, right slap, left thumb, and right thumb) captures and evaluated ten-finger rolled-to-rolled, plain-to-plain and plain-to-rolled identification with segmentation on database of 5 million records.
Another distinctive feature of the FpVTE12 tests is the searching time limit for one-to-many identification. Those participants who failed to meet time limits were dismissed.Thus all companies prepared their new high-speed identification algorithms. Not only searching speed but also the accuracy of identification was evaluated as well for final reports. The evaluation allowed each participant to make two submissions per class - fast and slower versions of algorithms.
There were 22 applicants for testing but only 18 could passat least one test. These 18 are mostly well-known companies in the world of biometrics including two Russian companies - Sonda and Papillon, and a number of new companies as well.
|C||Afis team||A, B, C|
|D||3M Cogent||A, B, C|
|E||Neurotechnology||A, B, C|
|F||Papillon||A, B, C|
|G||Dermalog||A, B, C|
|H||Hisign Bio-Info Institute||A, B, C|
|I||NEC||A, B, C|
|J||Sonda||A, B, C|
|L||Innovatrics||A, B, C|
|M||SPEX||A, B, C|
|O||ID Solutions||A, B, C|
|Q||Morpho||A, B, C|
|S||Decatur Industries||A, B, C|
|U||Aware||A, B, C|
|V||AA Technology||A, B, C|
Both experts and organizers of tests gave much attention to the results in Class C on database of 5 million records. The first three places on identification accuracy granted to three biometric giants: NEC (Japan), Morpho / Safran (France) and 3M Cogent (USA).
Sonda in the plain-to-rolled searching modetook the 4-th place right after the leaders, in plain-to-plain and rolled-to rolled modes Sonda was on the 5-th place.The error probability of 0.3% differed about two times from the group of leaders. But in some tests the searching speed and length of mathematical code of Sonda’s algorithms exceeded leaders. In all other tests Sonda invariably held positions in the top five. Thus Sonda's algorithms provided high accuracy while limiting time and computational resources for most finger combinations.
Following the testing results NIST marked out a group of five companies which submitted optimal algorithms based on three factors: the identification accuracy, searching speed and computational resources required. Those companies are: Sonda (Russia), Innovatrics (Slovakia), Morpho (France), IDSolutions (USA) and AA Technology (China).
Please visit the NIST site for the official report: http://dx.doi.org/10.6028/NIST.IR.8034