In the original Faster R-CNN paper, the authors parameterized the box coordinates for regression under RPN. Below is the snippet of how they computed it:

                                        enter image description here

Though this confuses me. They set the number of anchor boxes at 256. Though the number of ground-truth boxes (I'll call $n^*$) for each image is highly variable and is never equal to 256, unless by some coincidence. Same can be said for the number of predicted boxes ($n$). So how do they compute these numbers? Would there be $(4)(256)n$ predicted parameters and $(4)(256)n^*$ truth parameters? In the case that $n \neq n^*$, which is most definitely possible especially in early epochs, the loss cannot be computed as the lengths do not match.

What am I misunderstanding here? There must be something inferred from the article because the authors never touch on it again.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.