[tedmiston]

I would strongly guess it's a fixed set of product images they're training against, possibly attained by massive scraping. Another part of training or processing might consist of a reverse image search API, like TinEye, and gathering metadata from the pages containing the result images.