Forgive me if I'm misunderstanding how it works, is the model is predicting the original emoji from an augmented emoji, and doing that for every 36x36 patch in the webcam image?
edit: yep - predicting the original based on the augmented version (as a hack to generalize for unseen inputs)