Why perform rendering transformer by the entire feature map?

[coolbay]

May I ask why you perform rending transformer by the entire feature map instead of pixels? Does it work well if you do rendering transformer the same way for grounding transformer?

Thank you very much!

[dongzhang89]

@coolbay Thanks, very good question. There are two main reasons why we do this: 1. The transformer itself is a very computationally intensive operation. It is not necessary to use so many transformers in FPT; 2. It is actually meaningless to render a high-level object with the attributes of another distant low-level object or pixel positions.