Learning a fully convolutional net on minibatches of images of different sizes

Question

I am wondering if/how people are doing to learn a FCN on images that do not share the same dimensions. In theory, nothing prevents it, it is actually pretty straightforward.