Confusion about why a single conv layer is able to learn the offsets about dcnv2 HOT 5 OPEN

Can anyone explain? Thanks!

from dcnv2.

Hi, I think the offset is learned during the backpropagation. There is an update rule for the offset. You can take a look at equation 7 in the Deformable ConvNet (v1) paper.

from dcnv2.

@DKandrew @CharlesShang
Thanks for your reply!
And sorry that I didn't describe my question clearly. It doesn't seems to me to make sense that the learned offsets can point to the positions that outsize the receptive field of a conv kernel. For a set of offsets at a point in the input, since the conv operator is locally connected to the input features, I think it should only know where to point just within the receptive field of the kernel, but in the paper there is no constraint for this. I can't figure this out.

from dcnv2.

@eezywu @DKandrew @CharlesShang I am also confused.

from dcnv2.

@eezywu I think deform conv can only be used in higher layers which can provide larger receptive field. The offsets are actually influenced by previous layers.

from dcnv2.