[Question] Is it be possible to train Unet with RGB masks?

[michaelwachter]

Hey folks,

I'm pretty new to Machine Learning/Deep Learning. For my bachelor thesis I am trying to recreate this Paper with other Neural Networks. (Basically a classifier behind a fine tuned Autoencoder)

I know that Unet is meant to segment images. But I'm trying to reconstruct images in RGB with this Network.

I tried to feed the network RGB masks (which where just the original images) but I don't know how to modify the DataProvider to take output images with 3 channels.

Furthermore I would try to nullify the concat_and_crop connections, to prevent the network to just take the original image.

Any suggestions or better alternatives for this project?