My guess is that they trained the model for this with the data they collected from [Quickdraw](https://quickdraw.withgoogle.com/)