Fully convolutional networks need a crop layer to crop the shape of output like data layer.
implement in caffe: https://github.com/BVLC/caffe/blob/master/include/caffe/layers/crop_layer.hpp#L14