How does the joint loss work?

[tengteng95]

I wonder how the joint loss influence the predicted segmentation map?
As in your code, "mask" in used to get centers of pre-defined parts(line851 and line861~863). It seems that tf.argmax is not differentiable( cannot backpropagate gradients to "mask_prob").
So I wonder why it can help generate better segmentation map.
Thanks!