Confuse with Deep Image Homography Estimation

This paper introduce a way to generate training examples for measuring homography matrix, following are their solution

On step 4, they apply HBA on image A, optain A new Image B, then they crop the image B at positions p, which are same as the position of Image A.

I am confusing with step 4, Apply HBA on image B is mapping points of image B to the corresponding points of image A, but what is the meaning of applying HBA to image A?