For deep learning, can the size of the ground truth image and the predicted image be different? Let's say can the predicted image of size 64×64 can be compared with a ground truth of size 128×128?
I have been trying to run the U-net model with input size to the network as 64×64 and the predicted image also comes as 64×64, but the ground truth image is of size 128×128. Is that comparable for the given CNN? Or how should I compare it?
The purpose of the ground truth is to compute a loss function which takes as inputs the prediction and the ground truth to return a value such as the closer the prediction is from the ground truth the smaller is the returned value. So the real question is how are you going to design this loss function ?
It would be probably easier for you to start by just reducing the resolution of your ground truth dataset with PIL for instance.
The second step would be to change the architecture of the model to fit the resolution of your ground truth dataset if that's what you need.
Of course you also need to understand what loss function you will be using but it probably requires same resolution images.