Amazon trains retail systems with synthetic palm biometric database
Amazon has been steadily improving its biometric retail systems, according to a presentation by Vice President of Physical Retail and Technology Dilip Kumar at the company’s recent re:MARS 2022 event, with continuing technology development and synthetic data for algorithm training.
Kumar summed up his talk on Amazon’s Just Walk Out, Amazon One and Amazon Dash Cart technologies in a post to a company website.
Ongoing innovation in “sensors, optics, and machine vision algorithms” have enabled the company to reduce the number of cameras used in its Just Walk Out system, Kumar says.
Synthetic data was used to increase the accuracy of Amazon’s AI algorithms, such as for different lighting conditions in stores. Amazon One’s palm biometrics algorithms were also trained and tested to work for customers from different demographics, temperatures, and physical characteristics like calluses and wrinkles.
The potential and limitations of human pose estimation, a technology used in ‘cashierless’ shopping and other applications, is examined in a Forbes Technology Council post from Mobidev Founder and CEO Oleg Lola. Like Kumar, Lola identifies a lack of training data as a potential barrier to widespread adoption or such cutting-edge computer vision technologies, but again like Kumar, he sees increased availability of training data ushering in another leap forward in capabilities, in this case for pose estimation.
Kumar also discussed the encryption technologies Amazon uses to secure the various applications.
Choosing a biometric
Kumar discusses the selection of palm print and vein biometrics at some length in a video produced by the company.
The three criteria Amazon used to select its modality, which were that it must be contactless, it must be private, and it should involve an intentional, intuitive gesture.
“When you look at a palm, you can’t ascertain a person’s physical identity,” unlike face or voice, Kumar says on the second point.
In terms of the natural motion, Kumar says it is a close equivalent to people passing their phone over a surface, such as a QR code; an action people are already used to.
The misidentification rate so far is zero, according to Kumar. The company sets its algorithms to prioritize eliminating false positives over avoiding false negatives, accepting the need for backup forms of identification.
Amazon One also includes liveness detection algorithms.