Hi! I'm one of the authors. We'd be happy to answer any questions. :)
f(X,Y,Z) = (x,y) (Mapping from 3D to 2D using renderer)
g(x,y) = (u,v) (mapping xy and uv)
h(x,y) = (r,g,b) (actual rendered image)
p(u,v) = (r,g,b) (texture map image)
gradient from upper layer is going to be wrt h(x,y) - dL/dh(x,y). How do you propagate this to get dL/dp ?
(Please let me know if this is explained in detail somewhere. Thanks!)