So, presumably these masks were drawn manually. The three different colors in the input mask tell it to train 3 different models from the masked regions, and then the same three colors in the output mask tell it where to apply each trained model. So it trains a model for each apple and then applies that model to the next apple over. It's not just randomly swapping colors on the apples.
Input mask: https://github.com/luanfujun/deep-photo-styletransfer/blob/m... Output mask: https://github.com/luanfujun/deep-photo-styletransfer/blob/m...
So, presumably these masks were drawn manually. The three different colors in the input mask tell it to train 3 different models from the masked regions, and then the same three colors in the output mask tell it where to apply each trained model. So it trains a model for each apple and then applies that model to the next apple over. It's not just randomly swapping colors on the apples.