Cross-Cascading Architecture in DensePose

[nzqo]

I am trying to understand how to implement the cross-cascading architecture described in the original DensePose paper (Figure 8). I don't understand the "Combination" step specifically. The tensors to be combined seem to have incompatible dimensions, so a simple "+" as shown in the figure doesn't seem possible. I am also confused about the dimensions "[1/17] x 512 x 1 x 1", because of the 1/17 fraction. Does anyone have an idea?

Note there is an open unanswered issue basically asking the same thing I am in the now archived Densepose repository:
facebookresearch/DensePose#62