Amazon's Just Walk Out combines cameras, weight sensors and GAN-made training data

Amazon’s Just Walk Out technology, first opened to the public at an Amazon Go store in Seattle on January 22, 2018, decides who takes what so it can charge shoppers correctly when they leave. Amazon says it combines computer vision and machine learning with two main sensor types: ceiling-mounted cameras that give a store-wide view, and load-cell weight sensors on the shelves that help detect small items like gum or lipstick. The company calls this blend “sensor fusion.” To train the system to recognize shopping actions across different store layouts, lighting and crowds, Amazon generated synthetic training data, including using generative adversarial networks, so the models could learn from scenarios that would be hard to capture in real footage.