fcn-8s training and testing on different sized dataset

[avcs]

I want to fine tune pre-trained FCN-8s on my own dataset with four labels. Just read the "Fully Convolutional Networks for Semantic Segmentation" paper. one of the statements in the abstract says "Our key insight is to build “fully convolutional” networks that take input of arbitrary size and produce correspondingly-sized output with efficient inference and learning....."

Does this mean I could train fcn-8s on smaller image and test/ validate on large ones? That would be awesome because my images are fairly large and I could downsample it to create smaller labels. 

Thanks for the reply.