It's true, but since everybody has access to the test images (so they can label them), and the models (through google earth), it seems like anybody can train whatever algorithm they want on the same set.