ckkelvinchan / offset-fidelity-loss Goto Github PK

Dear Authors:
Thanks you very much for your excellent work. In the "Offset Diversity" section, you mention that "To verify their equivalence, we train two instantiations - original DCN and our decomposition". I would like to follow your work for new advancement, so could you release this code?
In addition, the visual results of the learned offsets such as Figure 4 are wonderful. How do you visualize this?

I am using the DCNv2 package released in github, I wonder how to get the learnable DCN offset for fidelity offset since the interface only provides the aligned output?
And is it still as efficient as using DCN when adding the modulated module since the modulated module has the similar function with fidelity loss?