Models are trained from scratch on small, visually coherent datasets. Set sizes have varied from 8 to a few thousand images, in some cases even more. Typical training times from a couple of hours to a few days. The goal is to learn visual features, not semantics.
