Why should we use two localization outputs?

[qjadud1994]

trafficstars

Hi, I have a question about structure of Faster R-CNN.

Why should we use two localization outputs from RPN and Classification Head?

I think only localization results from RPN is enough.

Can we proceed with class-specific classification only for positive boxes from RPN and just using that positive boxes coordinates for localization?

I'm so confused.

If anyone knows the answer, Please let me know.

[hushunda]

That like boosting

[chrisway613]

yep, if u only want to get the detected object, RPN is enough. But, in order to get the more precise results, it should combine with the detection head.