SAFR^® Inside: World-class face recognition, now on the edge

Accuracy

SAFR camera embedded face recognition achieves accuracy level of 99.87% on the publicly available LFW dataset and 0.0334 FNMR @ 0.0001 FMR on wild faces at NIST FRVT 1:1. It is measured by NIST to be the fastest and smallest algorithm out of top performing algorithms on wild faces. Network is also measured by NIST to have one of the least biased performance characteristics across skin tones and gender.

Metrics

Accuracy:  99.687 %

Precision: 99.897 %

Recall: 96.66 %

F1 Measure: 0.9825

NIST World-wide demographic uniformity rank: 2/103 (top 2%)

Above metrics are indicative of real-world algorithm performance on unconstrained face images with wide range of resolutions and wide yaw and pitch pose variations (wild images). The algorithm was measured on over 1 million positive matches and over 10 million negative (impostor) matches on population of over 10000 individuals. In above measurements, algorithm was configured to produce few face positives (1/10000) and thus above numbers do not represent maximum achievable accuracy but accuracy one would expect in practical deployments. The algorithm exhibits stable false match rates on increasing population sizes. Low false match rates, fast rate of operation and low memory requirements enable SAFR algorithm to be highly effective on video where multiple face images are available.

Algorithm also exhibits high performance uniformity across gender and skin tones. SAFR ranks 2nd (out of 103 algorithms) as least variable algorithm across gender and skin-tones according to evaluation by NIST on sampling of world-wide population.

For comparative evaluation, additional metrics are provided below. They reflect achieved accuracy on the LFW (Labeled Faces in the Wild) data set in unconstrained, labeled outside data category after applying known and documented error corrections to the benchmark data-set.

Accuracy:  99.867 %

Precision: 99.9 %

Recall: 99.833 %

F1 Measure: 0.98667