These classification networks are trained on ILSVRC for object recognition. We cast these nets into fully convolutional form to make use of their parameters as pre-training.
To reproduce our FCNs, or train your own on your own data, you need to first collect the corresponding base network.