I'm having trouble understanding something when going over Faster R-CNN paper (https://arxiv.org/pdf/1506.01497.pdf).

From what I understood, if k is the number of anchors per position on the feature map, RPN outputs a 4*k vector (for BB regression) and a 2*k vector (for obj/not obj classification) for each position on the feature map (so if i understood well the actual output is a 14x14x4*k and 14x14x2*k).

In the 3.1.3 Training RPNs section, it's said "we randomly sample 256 anchors in an image to compute the loss function of a mini-batch".

So my question is what are we suppose to do with anchors not used during training ? How not to take them into account while there is a value expected for them (class and reg)?


1 Answer 1


Just don't include them as part of the loss function -- since the 256 anchors are sampled randomly, eventually those anchors will probably show up later in the training.


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.