Amazon Web Services’ machine learning-based face recognition service correctly identified 90 percent of its subjects in tests run by analytics technology company Novetta using relatively small databases.
Novetta released its findings in late August. The analysis also found Amazon Rekognition enrollment error rates were about 10 times lower than those of other cloud-based face recognition tools and that Rekognition is calibrated to reduce false positive errors, even if that leads to relatively high false negative errors.
Rekognition launched in November 2016 and is being increasingly used for government and commercial purposes.
“Amazon Rekognition – along with other machine learning-based approaches – is emerging as a disruptive capability in U.S. government applications,” said Michael Thieme, Novetta’s vice president of special projects. “Understanding its real-world performance is a precondition of effective use in surveillance, large-scale identification, and social media applications.”
2 Comments
Rachel,
Your article misstates our analysis. You state that “The analysis also found Amazon Rekognition enrollment error rates were about 10 times higher than those of other cloud-based face recognition tools.” The error rates for Rekognition are, in fact, 1/10 of that of the other tools tested (.4% as compared to several over 4.8% and 5.4% for competitors). Please correct the article to read that that Rekognition error rates were 10x lower (or 1/10th) that of competitors we tested.
Thank you for the note, Myles! We have corrected the article.