Is there any code to distinguish foreground and background images?

[yuanyuan29]

As the author discribed in the chapter 4.5 in their TPAMI paper, they add another branch to do the classification for dataset including background images. But I am nor clear how they design the model so that for the bakcground images the gradients from the salient object detection module are not allowed to back-propagate. Is there any code related to this?