Questions about offset calculation

Hi, thanks for your wonderful work.

I’m currently studying your work, and come up with some question about the offset calculations.

I understood the offset calculation mentioned on the paper, but can’t understand about how generated offset is being used in the code.

For ex) if $S_H \times S_W : 3 \times 1$; I understood how the offset is applied in this figure

by calculate like this:

However, when I run the offset generating code, I can’t figure out how this offset is being used in deform_conv2d

Can you provide more detailed information about this??

And also, the paper contains how $S_H \times S_W: 3 \times 3$ works, but in the code, it seems like either one ofkernel_size[0] or kernel_size[1] has to be 1. So, if I want to use $S_H \times S_W : 3 \times 3$, do I have to make $3 \times 1$ and $1 \times 3$ offsets and add those together?

Thank you again for your work. I really learned a lot.

[tae-mo]

Thank you very much for your kind explanation! Have a great day◡̈

[ShoufaChen]

Yes, your understanding is correct.