why caffe softmaxwithloss layer can accept 2D label

[Di Jin]

I read from Caffe document that the label matrix size of the layer SoftmaxWithLoss is N*1*1*1, but in the train prototxt file of FCN paper (https://github.com/shelhamer/fcn.berkeleyvision.org/blob/master/voc-fcn32s/train.prototxt), they just use an image as the label whose size is N*1*H*W, so why is it that this kind of label size is acceptable? In addition, I want to implement the 2D softmaxwithloss layer like FCN but I need to the weight the loss pixel-wise, how to implement this? Thanks!