CityScapes of Data preprocessing

[Ricelll]

@phillipi when Training the model on CityScapes(photo->label), I have a question the how to prepare the pair of data(input,target): {RGB photo，the ont-hot label map},size is {(3,2048,1024),(30,2048,1024)} or {RGB photo，the ont-hot label map},size id {(3,2048,1024),(1,2048,1024)} or any else??

[junyanz]

In pix2pix and CycleGAN, we treat the label map as an RGB input (3 channels). It works fine for Cityscapes but might not work once you have many categories.