reference point visualization and some questions

[SnaKey0u0]

Hi author!

I would like to ask if you have tried the visualization of sampling location or reference point? Such as the picture in the appendix of the original paper of Deformable DETR

Because after passing the Linear layer of sampling_offset, the offset is unbound.
Therefore, the subsequent normalization steps such as offset normalize and 2 * sampling_location - 1 make me confused, because this does not guarantee that the range of interpolated sampling is reasonable.

I tried to visualize a trained model, but the distribution of sample points was very terrible.

Looking forward to your reply, thank you very much!