Brightics Visual Search Claims 6th Place in NIST FRVT “Face Mask Effects” Category

The face recognition algorithm of Brightics Visual Search, Samsung SDS’s visual analytics solution, took 6th place in the “Face Mask Effects” category of the latest Face Recognition Vendor Test (FRVT).

The National Institute of Standards and Technology (NIST), an organization under the United States Department of Commerce, tests and announces the accuracy and performance of face recognition vendor’s algorithms in various categories. Almost all major face recognition vendors from all over the world are currently participating in the FRVT.

In September, Samsung SDS submitted its algorithm to NIST FRVT, for the third time, in “1:1 Verification” and “Face Mask Effects” categories: 1) VisaBorder (comparison of a visa photo and a photo taken at immigration or emigration, 2) Mugshot +12 Years (comparison of a mugshot photo and one taken after at least 12 years), 3) Kiosk (comparison of kiosk face photos), 4) Face Mask Effects (comparison of a visa photo and a photo taken while wearing a face mask), etc.

In Face Mask Effects category, Samsung SDS’s algorithm was in top 6 with the False Non-Match Rate (@False Match Rate =0.00001) of 0.0127, a 28% improvement from 0.0176, the test result of the previous algorithm submitted to NIST FRVT in May.

In FRVT 1:1 Verification category, Samsung SDS ranked 5th for VisaBorder, the same rank it achieved in its last participation in May.

Samsung SDS and Samsung Advanced Institute of Technology (SAIT) have been cooperating to improve the face recognition algorithm since last year. They have continuously applied various new optimization technologies and training datasets and repeatedly tested the performances of the upgraded algorithms to get better results. The face dataset, made thanks to the voluntary participation of Samsung SDS colleagues, also played an important role in getting good results in the FRVT.

FRVT/FNMR Accuracy Ranking (Nov. 2021, May 2022, Sep. 2022)
Area and Category Nov. 2021 May 2022 Sep. 2022
Accuracy Ranking Accuracy Ranking Accuracy Ranking (change)
FRVT 1:1 Verification VISABORDER FNMR@FMR= 0.000001 0.0099 136 0.0024 5 0.0021 5
FRVT 1:1 Verification MUGSHOT (ΔT≥12YRS) FNMR@FMR= 0.00001 0.0081 132 0.0023 6 0.0022 5(△1)
FRVT 1:1 Verification KIOSK FNMR@FMR= 0.00001 0.0991 52 0.2598 156 0.0489 8 (△148)
FRVT FACE MASK Effects FNMR@FMR= 0.00001 - - 0.0176 12 0.0127 6 (△6)
* Notes: FNMR (False Non-Matching Rate) is the rate of instances where authorized persons are incorrectly classified as unauthorized persons. It is also called the FRR (False Reject Rate). FMR (False Matching Rate) is the rate of instances where unauthorized persons are incorrectly classified as the authorized persons. It is also called FAR (False Acceptance Rate). FNMR and FMR are always stated together when describing accuracy of a biometric system.

Until recently, Brightics Visual Search has been used mainly by media companies such as CJ ENM, for video metadata generation, advanced search, content recommendation and automated editing purposes. However, nowadays, more and more companies and industries are adopting this solution to analyze, search, edit and manage their video data and improve their visual communication services:

• Customer analysis in retail shops of global autobile and electronics manufacturers.
• Attendance check in online-education service of Brazilian local governments;
• Searching and editing of video manuals or promotional video materials in manufacturing companies and public organizations.

Samsung SDS’s customers are also adopting Brightics Visual Search along with RPA solution in business automation applications.