For the first time in the January 6th, 2020 report, the National Institute of Standards and Technology (NIST) Face Recognition Vendor Test (FRVT) “Ongoing” benchmark is reporting the peak memory usage of each face recognition algorithm. This measurement, which accounts for the RAM required to load an algorithm’s software libraries and statistical models, is a critical factor for developers of mobile, on-edge and embedded applications, as there is limited memory available for such applications.
In the January 21st, 2020 report, Rank One’s algorithms were 3rd and 5th in memory usage of the 190 benchmarked, using less than one tenth the memory of the median algorithm submission. The only algorithms to use less memory were extremely inaccurate (Rank One’s algorithms were over 50x more accurate than such submissions).
The ROC SDK v1.20 (listed as algorithm “rankone-008” in the NIST reports), used a mere 70MB of memory. By contrast, the median usage was 730MB, with several prominent algorithms requiring in excess of 1GB to perform a 1-to-1 identity verification. Such large memory requirements for these solutions render them unusable in most mobile and embedded system applications.
Rank One’s efficient memory usage, coupled with it’s industry leading template size that is 12x smaller than the median NIST algorithm template size, enrollment speeds that are over 2x faster than the median algorithm speed, and top-tier accuracy metrics, make Rank One a uniquely qualified solution for supporting mobile and “on-edge” applications.
Read our latest latest blog post. to learn more about the results of this benchmark, and why this efficiency metric and others are critical in the development of certain applications.