1
$\begingroup$

The authors of the original paper of Faster R-CNN when they refer to the positive anchors, they are labeled as 1. I guess they refer in binary classification. What happens in the case in a task we have more than two classes?

I mean the purpose of the RPN is to generate just region proposals (there is an object or not, regardless the label) or it generates region proposals and the respective label as well? I guess it is the second case, since the positive anchors that participate in the training process need to have a ground truth label as well, i.e. (cat, dog, e.t.c).

However, I might misunderstand things and interpret things in a wrong way.

$\endgroup$

1 Answer 1

1
$\begingroup$

As well as the algorithm used for region proposals in R-CNN and in Fast R-CNN, RPN task in Faster R-CNN is to propose regions that may or may not be an object, so its cls head with 2k scores corresponds to the classification of every anchor as an object or as a non object (this way all of them may be non objects): enter image description here

But it could also be a k scores logistic regression (k sigmoids 0 to 1) as they stated in the fourth note of the paper:

  1. For simplicity we implement the cls layer as a two-class softmax layer. Alternatively, one may use logistic regression to produce k scores

After proposal generation, Fast R-CNN is used to classify and improve every selected proposal, i.e. grey area in the following image:

enter image description here

So you may better conceptualize the difference between Fast and Faster R-CNN considering the region proposal generation algorithm as a black box, the one could be filled with a RPN head connected to the output convolutional feature map of the deep net used as backbone, or the standard Selective Search used by R-CNN and Fast R-CNN

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.