SAFR algorithm performance metrics are provided below. They reflect achieved accuracy using the LFW (Labeled Faces in the Wild) dataset in the “unconstrained labeled outside data” category after applying known and documented error corrections to the benchmark National Institute of Standards and Technology (NIST) worldwide demographic uniformity rank, These metrics mean that SAFR is ranked 2/103 in NIST, placing it in the top 2% of entrants.

Accuracy: 99.687% 

Precision: 99.897% 

Recall: 96.66% 

F1 Measure: 0.9825 

The above metrics are indicative of real-world algorithm performance on unconstrained face images with a wide range of resolutions and wide yaw, pitch, and roll pose variations. (i.e. wild images) The algorithm was measured on over 1 million positive matches and over 10 million negative (impostor) matches in a population of over 10,000 individuals. In the above measurements, the algorithm was configured to produce few face positives (1/10,000), and thus the above numbers do not represent maximum achievable accuracy but rather the accuracy one would expect in practical deployments. The algorithm exhibits stable false-match rates with increasing population sizes. Low false-match rates, a fast rate of operation, and low memory requirements enable the SAFR algorithm to be highly effective with videos where multiple face images are visible. 

The SAFR facial recognition algorithm is optimized for wild-face recognition in videos in real time, maximizing recognition accuracy while retaining a high speed of operation. It's built on a state-of-the-art algorithm developed using artificial intelligence, which ensures future readiness as well as the best accuracy and speed. Among algorithms with low false non-match rates (FNMR) for wild faces, SAFR is the fastest and smallest — 3.7x faster and 7.14x smaller than the average algorithm. This combination of speed and accuracy makes SAFR the highest effective accuracy algorithm for live video and thus real-time operations.


For more information see The Fastest Facial Algorithm Just Got Faster.